Carolina Neves Souza

Tese_Carolina_Neves_Souza_2024.pdf
Documento PDF (2.5MB)

UNIVERSIDADE FEDERAL DE ALAGOAS
INSTITUTO DE CIÊNCIAS BIOLÓGICAS E DA SAÚDE
Programa de Pós-Graduação em Diversidade Biológica e Conservação nos
Trópicos

CAROLINA NEVES SOUZA

A CONTRIBUIÇÃO DAS REDES SOCIAIS PARA A CONSERVAÇÃO:
compreendendo os interesses e percepções digitais sobre as áreas protegidas
brasileiras.

MACEIÓ - ALAGOAS
Março/2024

CAROLINA NEVES SOUZA

A CONTRIBUIÇÃO DAS REDES SOCIAIS PARA A CONSERVAÇÃO:
compreendendo os interesses e percepções digitais sobre as áreas protegidas
brasileiras.

Tese apresentada ao Programa de Pós-Graduação em
Diversidade Biológica e Conservação nos Trópicos,
Instituto de Ciências Biológicas e da Saúde.
Universidade Federal de Alagoas, como requisito para
obtenção do título de Doutora em CIÊNCIAS
BIOLÓGICAS, área de concentração em Conservação
da Biodiversidade Tropical.

Orientadora: Profa. Dra. Ana Cláudia M. Malhado
Coorientador: Prof. Dr. Ricardo A. Correia
Coorientadora: Profa. Dra. Adriana R. Carvalho

MACEIÓ - ALAGOAS
Março/2024

Catalogação na Fonte
Universidade Federal de Alagoas
Biblioteca Central
Divisão de Tratamento Técnico
Bibliotecário: Marcelino de Carvalho Freitas Neto – CRB-4 - 1767
S729c

Souza, Carolina Neves.
A contribuição das redes sociais para a conservação : compreendendo os interesses
e percepções digitais sobre as áreas protegidas brasileiras / Carolina Neves Souza. –
2024.
114 f. : il.
Orientadora: Ana Cláudia M. Malhado.
Co-orientador: Ricardo A. Correia.
Co-orientadora: Adriana R. Carvalho.
Tese (Doutorado em Ciências Biológicas) – Universidade Federal de Alagoas.
Instituto de Ciências Biológicas e da Saúde. Programa de Pós-Graduação em
Diversidade Biológica e Conservação nos Trópicos. Maceió, 2024.
Inclui bibliografias.
Apêndices: f.107-114.
1. Twitter (Rede social on-line) - Engajamento. 2. Análise de sentimento. 3.
Monitoramento ambiental. 4. Dados digitais. 5. Culturômica. I. Título.
CDU: 504(81)

Folha de aprovação

Carolina Neves Souza
A CONTRIBUIÇÃO DAS REDES SOCIAIS PARA A CONSERVAÇÃO:
compreendendo os interesses e percepções digitais sobre as áreas protegidas
brasileiras
Tese apresentada ao Programa de PósGraduação em Diversidade Biológica e
Conservação nos Trópicos, Instituto de
Ciências
Biológicas
e
da
Saúde.
Universidade Federal de Alagoas, como
requisito para obtenção do título de Doutor(a)
em CIÊNCIAS BIOLÓGICAS na área da
Biodiversidade.

Tese aprovada em 05 de março de 2024.

Dr.(a) Ana Claudia Mendes Malhado/UFAL
(orientadora)

Dra. Adriana Rosa Carvalho
(Coorientadora)

Dr. Ricardo Aleixo Correia
(Coorientador)

Dr. (a) Alexandre Schiavetti

Dr. (a) Rafael Ricardo Vasconcelos da Silva
Dr. (a) Marta de Azevedo Irving

Dr. (a) Guilherme Ramos Demétrio Ferreira
MACEIÓ – AL
Março/ 2024

À minha filha Clarice, que me inspira todos os dias
a conservar e ser mais curiosa sobre nossa
relação com a natureza.

AGRADECIMENTOS

Gostaria de expressar minha gratidão à minha orientadora, Profa. Dra. Cláudia
Mendes Malhado pela sua orientação, amizade e apoio ao longo desta jornada
acadêmica.

Seu

comprometimento

paciência

foram

fundamentais

para

desenvolvimento desta tese de doutorado. Além disso, agradeço aos meus
coorientadores, Prof. Dr. Ricardo Correia e Profa. Dra. Adriana Carvalho por suas
valiosas contribuições e ideias que enriqueceram significativamente a qualidade deste
trabalho. Gostaria ainda de agradecer ao Dr. Ricardo por toda a sua generosidade e
paciência ao ensinar e compartilhar seu conhecimento e ideias.
Estendo meus agradecimentos aos pesquisadores que me acolheram durante o
período de doutorado-sanduíche na Universidade do Porto em Portugal, Msc. Javier
Martinez e Dra. Ana Sofia Vaz. Sua hospitalidade foi muito importante para o meu
crescimento acadêmico. Um abraço especial aos colegas do LACOS21, cuja colaboração
e amizade foram essenciais para enfrentar os desafios e celebrar os sucessos,
especialmente ao Prof. Dr. Richard Ladle, João Arthur Almeida, Thainá Lessa, Evelynne
Cardoso, Ludmila Costa-Pinto, Ana Carla Rodrigues, Jhonatan Guedes-Santos e Felipe
Vieira. Finalmente, reconheço o trabalho dedicado e as contribuições construtivas dos
revisores dos artigos científicos originados deste trabalho e dos membros das minhas
bancas de avaliação do doutorado, cujas sugestões valiosas contribuíram para a
qualidade e rigor deste estudo. A todos, meu sincero agradecimento por fazerem parte
desta jornada científica.
Agradeço ao CNPq, ao FUNBIO e ao Instituto Humanize pelas bolsas concedidas.
Os apoios financeiros destas instituições foram vitais para realização desta pesquisa, e
para que eu, enquanto pesquisadora, pudesse me dedicar integralmente a este trabalho.
Por fim, para além do ambiente acadêmico, desejo expressar minha profunda
gratidão à minha família, que sempre me apoiou, incentivou meus estudos e acreditou
nos meus sonhos. Um agradecimento especial é destinado ao meu marido, Iran, e a
minha filha, Clarice, pelo amor e compreensão que tornaram esta jornada possível.

RESUMO

As áreas protegidas (APs) desempenham um papel vital na proteção do patrimônio
natural e cultural, ao mesmo tempo que apoiam os meios de subsistência locais. No
entanto, apesar dessa importância, no Brasil, elas enfrentam desafios ligados à falta de
recursos financeiros e a percepção de baixa eficácia da gestão, o que pode resultar na
falta de apoio da sociedade com relação a essas áreas. Diante desse cenário, é
necessário compreender melhor as percepções e interesses do público pelas APs
brasileiras. Nesta tese, foram utilizados dados da rede social Twitter (renomeada X) para
explorar tendências de interesse e percepções sobre as áreas protegidas no Brasil. Para
isso, na primeira fase, foram coletados tweets em português sobre todas as categorias
de APs no período de 2011 a 2020. Em seguida, uma metodologia inovadora de análise
de sentimentos foi aplicada, focando especificamente nos parques nacionais (PARNAs)
brasileiros, e ampliando o período de coleta de dados até 2022. Os conteúdos textuais
dos tweets foram analisados com base em métricas de postagem e engajamento dos
usuários (curtidas e retuítes), classificação dos sentimentos expressos no texto e
modelagem de tópicos. Os resultados indicam que o número de usuários/tweets que
postam sobre as APs brasileiras permaneceu estável ao longo do período amostral, no
entanto, o engajamento cresceu consideravelmente a partir de 2018, coincidindo com
mudanças no governo federal. Embora os parques nacionais tenham recebido mais
menções, especialmente relacionadas as atividades turísticas, os tweets relacionados à
conflitos nas APs atraíram mais discussões. A análise de sentimentos identificou 18.388
(17,30%) de postagens expressando sentimentos negativos em relação aos PARNAs,
sendo a maioria relacionada aos incêndios florestais ocorridos entre 2011 e 2017 e ao
impacto das decisões governamentais que afetaram os esforços de conservação pós2019. Foram identificados seis tópicos de discussão proeminentes: (1) Incêndios
florestais; (2) Segurança; (3) Regramentos; (4) Vida selvagem morta por atropelamento;
(5) Privatização (Concessões); e, (6) Falta de recursos financeiros, refletindo a variedade
de sentimentos negativos em relação aos parques. Além disso, a modelagem de tópicos
por parques revelou-se benéfica na identificação de diferentes problemas e conflitos nos
cinco PARNAs mais tuitados, facilitando assim ações de conservação direcionadas.
Neste sentido, o estudo destaca a importância da análise de dados das redes sociais
para compreender o interesse público e promover uma gestão mais eficaz das áreas
protegidas. Isso pode subsidiar ações de conservação, melhorar a experiência dos
visitantes e comunicar a importância das APs para a sociedade. Em última análise os
resultados ressaltam o valor da culturômica para identificar lacunas e promover melhorias
que alcancem o apoio público as áreas protegidas brasileira.
Palavras-chave: Engajamento. Análise de sentimento. Twitter. Monitoramento
ambiental. Dados digitais. Culturômica.

ABSTRACT

Protected areas (PAs) play a vital role in protecting natural and cultural heritage while
supporting local livelihoods. However, despite this importance, in Brazil they face
challenges linked to a lack of financial resources and the perceived low effectiveness of
management, which can result in a lack of support from society for these areas. Given
this scenario, it is necessary to better understand the public's perceptions and interests in
Brazilian PAs. In this thesis, data from the social media Twitter (renamed X) was used to
explore trends in interest and perceptions of protected areas in Brazil. To do this, in the
first phase, tweets were collected in Portuguese about all categories of PAs from 2011 to
2020. Next, an innovative sentiment analysis methodology was applied, focussing
specifically on Brazilian national parks (PARNAs) and extending the data collection period
to 2022. The textual content of the tweets was analysed based on posting metrics and
user engagement (likes and retweets), classification of the sentiments expressed in the
text and topic modelling. The results indicate that the number of users/tweets posting
about Brazilian PAs remained stable throughout the sample period, however, engagement
grew considerably from 2018 onwards, coinciding with changes in the federal government.
Although national parks received more mentions, especially related to tourism activities,
tweets related to conflicts in PAs attracted more discussion. Sentiment analysis identified
18,388 (17. 30%) posts expressing negative sentiment towards PARNAs, with the majority
related to the forest fires that occurred between 2011 and 2017 and the impact of
government decisions affecting post-2019 conservation efforts. Six prominent discussion
topics were identified: (1) Forest fires; (2) Security; (3) Regulations; (4) Wildlife killed by
trampling; (5) Privatisation (Concessions); and, (6) Lack of financial resources, reflecting
the variety of negative sentiments towards the parks. In addition, topic modelling by park
proved beneficial in identifying different problems and conflicts in the five most tweeted
PARNAs, thus facilitating targeted conservation actions. In this sense, the study highlights
the importance of analysing social media data to understand public interest and promote
more effective management of protected areas. This can subsidise conservation actions,
improve the visitor experience and communicate the importance of PAs to society.
Ultimately, the results emphasise the value of culturomics in identifying gaps and
promoting improvements that achieve public support for Brazilian protected areas.
Keywords: Engagement. Sentiment analysis. Twitter. Environmental monitoring. Digital
data. Culturomics.

LISTA DE FIGURAS

Revisão de literatura
Figura 1 - Boxplot representando a distribuição do volume de pesquisa relativa global 25

Capítulo 1
Figura 1 - Methodological flowchart................................................................................ 41
Figura 2 - Volume of posts, users and engagement of Brazilian protected areas .......... 46
Figura 3 - Bootstrapped estimates of mean post engagement per year ......................... 47
Figura 4 - Public interest related to the Brazilian protected areas over the years........... 48
Figura 5 - Bootstrapped estimates of mean post engagement per year and per user type
....................................................................................................................................... 49
Figura 6 - Geographical distribution of the tweets posted about Brazilian PAs .............. 50
Figura 7 - Brazilian PAs that have generated the most public interest and engagement 51
Figura 8 - Most published content related to Brazilian protected areas .......................... 54

Capítulo 2
Figura 1 - Map of distribution of 74 Brazilian national parks ........................................... 76
Figura 2 - Methodological flowchart................................................................................ 77
Figura 3 - The daily counts related to non-negative sentiments of Twitter posts regarding
Brazilian national parks from January (2011) to December 2022 ................................... 84
Figura 4 - The daily counts related to negative sentiments of Twitter posts regarding
Brazilian national parks from January (2011) to December 2022 ................................... 85
Figura 5 - A bar chart representing the dominant negative topics per park .................... 86

Apêndice B
Figura 1 - Word frequency about product reviews on Buscapé .................................... 110
Figura 2 - Boxplot illustrating the sentiment per word in the reviews from the Buscapé
website ......................................................................................................................... 110

LISTA DE TABELAS

Capítulo 2
Tabela 1 - Topics identified by the BERTopic models during the analysis of the negative
tweet dataset. Models 1 to 4 were performed with the nr_topics to auto and different min
cluster sizes for the HDBSCAN parameter model. ......................................................... 87

Apêndice A
Tabela 1 - Keywords used in the query for tweets about Brazilian protected areas ..... 107

11
SUMÁRIO1

1 Apresentação............................................................................................................. 13
Referências ................................................................................................................... 15
2 Revisão da literatura ................................................................................................. 17
2.1 Áreas protegidas brasileiras ................................................................................. 17
2.2 Percepções, atitudes e sentimentos relacionados as áreas protegidas ........... 20
2.3 Redes sociais e conservação ambiental .............................................................. 22
2.3.1 Limitações das redes sociais como ferramenta investigativa ................................ 26
Referências ................................................................................................................... 27
3 Objetivos .................................................................................................................... 34
3.1 Objetivo geral ......................................................................................................... 34
3.2 Objetivos específicos............................................................................................. 34
4 Capítulo 1 - Assessing Brazilian protected areas through social media: insights
from 10 years of public interest and engagement ..................................................... 35
4.1 Abstract ................................................................................................................... 36
4.2 Introduction ............................................................................................................ 36
4.3 Material and Methods ............................................................................................. 39
4.3.1 Brazilian protected areas ....................................................................................... 39
4.3.2 Data collection ....................................................................................................... 40
4.3.3 Data analysis ......................................................................................................... 43
4.4 Results .................................................................................................................... 46
4.4.1 Volume of tweets about Brazilian PAs ................................................................... 46
4.4.2 Users characteristics ............................................................................................. 48
4.4.3 Content analysis of tweets..................................................................................... 50
4.4.4 Which protected areas generate most public interest and engagement? .............. 51

Este sumário está organizado conforme a estrutura da tese, a qual foi dividida em dois capítulos
elaborados no formato de artigos científicos. Os artigos e suas respectivas seções foram escritos na língua
inglesa, de acordo com as exigências de submissão às revistas científicas pertinentes.

12
4.4.5 Geographic focus of tweets ................................................................................... 53
4.5 Discussion .............................................................................................................. 55
4.6 Conclusions ............................................................................................................ 60
4.7 Acknowledgments .................................................................................................. 61
References .................................................................................................................... 61
5 Capítulo 2 - Using social media and machine learning to understand negative
sentiments towards Brazilian National Parks ........................................................... 69
5.1 Abstract ................................................................................................................... 70
5.2 Introduction ............................................................................................................ 70
5.3 Material and Methods ............................................................................................. 74
5.3.1 Study area ............................................................................................................. 75
5.3.2 Data collection ....................................................................................................... 76
5.3.3 Sentiment analysis ................................................................................................ 78
5.3.4 Time series analysis .............................................................................................. 80
5.3.5 Topic modelling analysis ....................................................................................... 81
5.4 Results .................................................................................................................... 82
5.4.1 Public perceptions about Brazilian National Parks ................................................ 82
5.4.2 Trends of non-negative perceptions over time ...................................................... 83
5.4.3 Trends of negative perceptions over time.............................................................. 84
5.4.4 Main topics associated negative perceptions ........................................................ 85
5.5 Discussion .............................................................................................................. 89
5.5.1 Non-negative perceptions of Brazilian National parks ........................................... 89
5.5.2 Negative perceptions of Brazilian National parks .................................................. 90
5.5.3 Potential, limitations and future research............................................................... 93
References .................................................................................................................... 95
6 Considerações finais .............................................................................................. 103
Apêndice A - Palavras-chave utilizadas na coleta de dados no Twitter ..................... 107
Apêndice B - Informações metodológicas sobre a análise de sentimento com dados de
tweets em português .................................................................................................... 108

13
1 APRESENTAÇÃO

As áreas protegidas (APs) são essenciais para a conservação da biodiversidade e
possuem significativo valor cultural e socioeconômico no Brasil (Maretti et al., 2012). O
sistema brasileiro de APs é um dos maiores do mundo, mas enfrenta uma série de
desafios que ameaçam sua integridade e sustentabilidade em longo prazo. Entre os
desafios estão a dependência de sistemas de governança ultrapassados; crise
institucional e falta de transparência (Gerhardinger et al., 2011; Bragagnolo et al. 2016).
Além disso, as APs podem ser enxergadas por parte dos cidadãos e formuladores de
políticas como obstáculos ao desenvolvimento socioeconômico (Bernard et al., 2014).
Estudos na área destacam a falta de apoio da sociedade como uma das principais
barreiras para a efetividade das APs (Souza, 2017; Cozzolino e Irving, 2016; McClanahan
et al., 2005), ressaltando a necessidade de uma abordagem inclusiva que promova uma
boa comunicação entre o órgão gestor das áreas protegidas e seus usuários (Macedo e
Medeiros, 2018).
As interações humanas com as áreas protegidas variam amplamente, oferecendo
benefícios físicos e psicológicos. No entanto, as restrições de uso do espaço e a falta de
conhecimento sobre a importância dessas áreas, podem contribuir para o desinteresse
público e gerar um sentimento de distanciamento e desconexão das pessoas em relação
ao meio ambiente natural (de Hann et al., 2014). O apoio público desempenha um papel
fundamental na legitimação das APs, por isso é importante compreender essas
interações, sentimentos e interesses. Neste contexto, a coleta e análise de percepções
online através das redes sociais ou outras plataformas surge como uma ferramenta
promissora para alcançar esse entendimento (Ladle et al., 2017). Estes métodos têm sido
utilizados com sucesso para avaliar o ecoturismo em ambientes conflituosos, aumentar
a sensibilização pública sobre a conservação e compreender as percepções e
sentimentos do público sobre as áreas protegidas (Toivonen et al., 2019). No Brasil,
dados do Google Trends e Wikipedia já foram utilizados anteriormente para avaliar o
interesse público e a importância da Internet em relação às APs brasileiras (GuedesSantos et al., 2021; Correia et. al 2018). No entanto, pesquisas adicionais são

14
imprescindíveis para uma análise mais aprofundada da interação entre os usuários das
redes sociais e as áreas protegidas. É importante investigar não apenas o conteúdo
publicado pelos usuários, mas também sua origem geográfica, os tópicos mais debatidos
e compartilhados, além de uma análise mais detalhada dos sentimentos do público em
relação a essas áreas.
Nesta tese, foram utilizados dados da plataforma da rede social Twitter
(recentemente renomeada X) para investigar o interesse e os sentimentos do público
sobre as APs brasileiras. O Twitter é uma plataforma de rede social e microblogging
bastante popular, com mais de 666 milhões de usuários ativos em 2023 em todo o mundo
(Statista, 2023a). No Brasil, o Twitter tem em torno de 24,3 milhões de usuários ativos
(Statista, 2023b), que publicam milhões de comentários (os chamados "tweets") todos os
dias contendo pensamentos e opiniões. O Twitter é muito usado por jornalistas,
cientistas, políticos, gestores e pela sociedade em geral (Collins et al. 2016; Mohammadi
et al. 2018) para divulgar informações, promover o discurso público e, portanto, serve
como um barômetro potencialmente sensível da opinião pública.
Desta maneira, esta tese representa uma oportunidade significativa para ampliar
a discussão sobre o interesse da sociedade em relação às APs brasileiras. Isso envolve
não apenas a expansão do diálogo prático, mas também do debate teórico, relacionado
as abordagens metodológicas inovadoras, como a aplicação da análise de dados de
redes sociais para compreender as percepções e sentimentos online sobre as áreas
protegidas no Brasil.
A fim de abordar os diversos aspectos desta discussão, esta tese foi dividida em
dois capítulos elaborados em formato de artigo científico. O primeiro capítulo intitulado:
Assessing Brazilian protected areas through social media: insights from 10 years of public
interest and engagement, aborda a análise das métricas de publicação e engajamento
dos usuários do Twitter (curtidas e retuítes), buscando identificar padrões e fatores
relacionados ao interesse público em todas as áreas protegidas brasileiras. O segundo
capítulo intitulado: Using social media and machine learning to understand negative
sentiments towards Brazilian National Parks, concentra-se na análise dos sentimentos da

15
sociedade em relação aos parques nacionais brasileiros e explora como essa ferramenta
analítica pode contribuir para a identificação de problemas e conflitos nestas áreas.

REFERÊNCIAS
Bernard, E., Penna, L. A. O., & Araújo, E. (2014). Downgrading, downsizing,
degazettement, and reclassification of protected areas in Brazil. Conservation Biology,
28(4), 939–950. https://doi.org/10.1111/cobi.12298
Bragagnolo, C., Costa Gamarra, N., Claudia Mendes Malhado, A., & James Ladle, R.
(2016). Proposta Metodológica para Padronização dos Estudos de Atitudes em
Comunidades Adjacentes às Unidades de Conservação de Proteção Integral no Brasil.
Biodiversidade Brasileira, 6(1), 190–208.
Collins, K., Shiffman, D., & Rock, J. (2016). How are scientists using social media in the
workplace? PLoS ONE, 11(10). https://doi.org/10.1371/journal.pone.0162680
Correia, R. A., Jepson, P., Malhado, A. C. M., & Ladle, R. J. (2018). Culturomic
assessment of Brazilian protected areas: Exploring a novel index of protected area
visibility. Ecological Indicators, 85, 165–171.
https://doi.org/10.1016/j.ecolind.2017.10.033
Cozzolino, L. F. F., & Irving, M. de A. (2016). Por Uma Concepção Democrática De
Governança Para a Esfera Pública. Revista Políticas Públicas, 19(2), 497.
https://doi.org/10.18764/2178-2865.v19n2p497-508
De Haan, F. J., Ferguson, B. C., Adamowicz, R. C., Johnstone, P., Brown, R. R., &
Wong, T. H. F. (2014). The needs of society: A new understanding of transitions,
sustainability and liveability. Technological Forecasting and Social Change, 85, 121–
132. https://doi.org/10.1016/j.techfore.2013.09.005
Gerhardinger, L. C., Godoy, E. A. S., Jones, P. J. S., Sales, G., & Ferreira, B. P. (2011).
Marine protected dramas: The flaws of the Brazilian national system of marine protected
areas. Environmental Management, 47(4), 630–643. https://doi.org/10.1007/s00267010-9554-7
Ladle, R. J., Jepson, P., Correia, R. A., & Malhado, A. C. M. (2017). The power and the
promise of culturomics. Frontiers in Ecology and the Environment, 15(6), 290–291.
https://doi.org/10.1002/fee.1506
Lockwood, M. (2010). Good governance for terrestrial protected areas: A framework,
principles and performance outcomes. Journal of Environmental Management, 91(3),
754–766. https://doi.org/10.1016/j.jenvman.2009.10.005

16
Macedo, H. S., & Medeiros, R. P. (2018). Rethinking governance in a Brazilian multipleuse marine protected area. Marine Policy, July, 0–1.
https://doi.org/10.1016/j.marpol.2018.08.019
Maretti, Cláudio C; Catapan, Marisete; Abreu, Maria Jasylene;Oliveira, J. E. Dantas.
(2012). Áreas protegidas: definições, tipos e conjuntos – reflexões conceituais e
diretrizes para a gestão.
McClanahan, T., Davies, J., & Maina, J. (2005). Factors influencing resource users and
managers’ perceptions towards marine protected area management in Kenya.
Environmental Conservation, 32(1), 42–49. https://doi.org/10.1017/S0376892904001791
Mohammadi, E., Thelwall, M., Kwasny, M., & Holmes, K. L. (2018). Academic
information on Twitter: A user survey. In PLoS ONE (Vol. 13, Issue 5). Public Library of
Science. https://doi.org/10.1371/journal.pone.0197265
Souza, C. N., de Barros, E. L. S. F. C., Dantas, I. F. V., Bragagnolo, C., Malhado, A. C.
M., & Selva, V. F. (2022). Inclusion and governance in the managing Council of the
Costa dos Corais Environmental Protection Area. Ambiente e Sociedade, 25.
https://doi.org/10.1590/1809-4422ASOC20210074R1VU2022L3AO
Statista. Most popular social networks worldwide as of October 2023, ranked by number
of monthly active users. 2023a. Disponível em: <
https://www.statista.com/statistics/272014/global-social-networks-ranked-by-number-ofusers/> Acessado em: 15 de janeiro de 2024.
Statista. Leading countries based on number of X (formerly Twitter) users as of January
2023b. Disponível em: <https://www.statista.com/statistics/242606/number-of-activetwitter-users-in-selected-countries/> Acessado em: 15 de janeiro de 2024.
Toivonen, T., Heikinheimo, V., Fink, C., Hausmann, A., Hiippala, T., Järv, O., Tenkanen,
H., & Di Minin, E. (2019). Social media data for conservation science: A methodological
overview. Biological Conservation, 233(January), 298–315.
https://doi.org/10.1016/j.biocon.2019.01.023

17
2 REVISÃO DA LITERATURA

A presente revisão de literatura está dividida em três partes, as quais
fundamentam a pesquisa desta tese. A primeira parte introduz o conceito e os referenciais
teóricos, proporcionando uma visão abrangente dos principais desafios para a efetividade
das áreas protegidas brasileiras. A segunda parte destaca os conceitos relacionados à
percepção ambiental subjacentes a este estudo e explora como as teorias
comportamentais podem ser aplicadas ao ambiente das redes sociais. Por fim, a terceira
parte apresenta tanto o potencial quanto as limitações do uso das redes sociais como
ferramenta de investigação das percepções e sentimentos da sociedade em relação à
conservação da natureza.
2.1 Áreas protegidas brasileiras
As Áreas Protegidas (APs) desempenham um papel crucial como estratégia para
fomentar a preservação dos recursos biológicos e a utilização sustentável dos benefícios
naturais, abrangendo serviços ecossistêmicos e práticas culturais em escala global
(Maretti, et al., 2012; Watson et al., 2014). No Brasil, as áreas protegidas são legalmente
instituídas como unidades de conservação2. De acordo com a legislação que institui o
Sistema Nacional de Unidades de Conservação (SNUC), uma unidade de conservação
é definida como um “espaço territorial e seus recursos ambientais, incluindo as águas
jurisdicionais, com características naturais relevantes, legalmente instituído pelo Poder
Público, com objetivos de conservação e limites definidos, sob regime especial de
administração, ao qual se aplicam garantias adequadas de proteção” (BRASIL, 2000, art.
2º, inciso I). As terras indígenas também contribuem significativamente para a
conservação da biodiversidade, embora não sejam formalmente reconhecidas como
unidades de conservação pelo SNUC.

Vale ressaltar que, a presente pesquisa concentrou-se exclusivamente nas unidades de conservação. No entanto, a
fim de manter a consistência terminológica em toda a tese, incluindo os artigos publicados nas revistas apresentadas
nos próximos capítulos, as unidades de conservação são consideradas, neste trabalho, como áreas protegidas.

18
Dada a condição de país megadiverso, o Brasil possui uma grande
responsabilidade na conservação de sua biodiversidade (Rylands e Brandon, 2005).
Segundo o Cadastro Nacional de Unidades de Conservação (CNUC) de 2023, o Brasil
abriga um total de 2.859 áreas protegidas, que cobrem coletivamente uma área de
aproximadamente 2.583.237 km2. Considerando a área continental brasileira, 19,01%
estão protegidos por APs; enquanto na zona econômica exclusiva brasileira, 26,49%
estão sob proteção por APs (MMA, 2023).
O estabelecimento de um sistema nacional de áreas protegidas e a ampliação da
criação destas áreas, está em consonância com as metas globais de conservação
delineadas pela Convenção sobre Diversidade Biológica (CDB, 1992; BRASIL, 1994;
BRASIL, 1998). Entretanto, permanece a incerteza quanto à efetividade dessas áreas na
conservação da biodiversidade e em seus impactos socioeconômicos no Brasil. As áreas
protegidas tiveram origem no modelo americano de Yellowstone, o primeiro Parque do
Mundo criado em 1872. A adoção desse modelo, estimulou a exclusão das populações
tradicionais residentes dentro ou nos arredores dessas áreas as quais foram
estabelecidas primordialmente para a preservação dos recursos naturais e para servirem
como locais para a apreciação da natureza (DIEGUES, 1996). Essa abordagem
preservacionista, desde o início, restringiu a capacidade das populações tradicionais de
se identificarem com as áreas protegidas e de desenvolverem um senso de
pertencimento e responsabilidade sobre elas.
Em decorrência disso, diversos episódios de degradação ambiental foram
desencadeados pela emergência de conflitos resultantes da marginalização das
comunidades tradicionais presentes nessas áreas. Paralelamente, a adoção de práticas
não sustentáveis e o avanço urbano, aliado aos padrões crescentes de consumo,
contribuíram para a proliferação de novas categorias de áreas protegidas em todas as
esferas governamentais (SELVA et al., 2016). Os desafios enfrentados pelo sistema
brasileiro de áreas protegidas, além daqueles decorrentes do modelo top-down de sua
criação, são notáveis. Este sistema, um dos mais vastos do mundo, confronta ameaças
que comprometem sua integridade e sustentabilidade em longo prazo. Dentre tais
desafios, destacam-se:: i) a dependência de sistemas de governança desatualizados e
centralizados, prejudicando a participação direta da sociedade (Engen et al., 2021); ii)

19
uma crise institucional persistente e de longa duração na agência federal responsável
pela conservação da biodiversidade (Gerhardinger et al., 2011); iii) a falta de
transparência nas práticas de gestão e na comunicação sobre a importância dessas
áreas, intensificando a distância entre a sociedade e a natureza e podendo resultar no
aumento dos crimes ambientais (Bragagnolo et al., 2016); e, iv) um déficit crescente de
financiamento, enfraquecendo as áreas protegidas que não conseguem cobrir
integralmente seus custos de gestão (Silva et al., 2021).
Ademais, as áreas protegidas são frequentemente interpretadas como custos de
oportunidade, representando obstáculos ao desenvolvimento econômico segundo
perspectivas de alguns políticos e tomadores de decisão (Ferreira et al., 2014). De fato,
eventos de recategorização, redução e extinção de áreas protegidas (PADDD)
impactaram uma extensão de 72.892 km2 de áreas protegidas no Brasil entre os anos de
1981 e 2012 (Bernard et al., 2014). Salvaguardar as áreas protegidas brasileiras contra
os eventos PADDD exige que as atitudes negativas sejam combatidas por meio da
demonstração de seu valor para a sociedade, conforme exemplificado por Jepson et al.
(2017), e da evidenciação junto aos políticos de que tais áreas contam com um amplo
respaldo público (Guedes-Santos et al., 2021). Em síntese, é de suma importância
fomentar uma conexão mais robusta entre as áreas protegidas brasileiras e a sociedade,
com vistas a evitar que sejam percebidas como custos de oportunidade pelos cidadãos
e líderes políticos (Bernard et al., 2014).
A compreensão da percepção das pessoas sobre o ambiente protegido não
apenas é crucial para avaliar o valor das áreas protegidas para a sociedade, mas também
desempenha um papel fundamental na eficácia dessas áreas. A análise conduzida por
Macedo e Medeiros (2018) destaca que a colaboração da sociedade na promoção da
eficácia das áreas protegidas está intrinsecamente vinculada à interação entre os
incentivos participativos e os incentivos ao conhecimento. Nesse sentido, a promoção de
ações transparentes e comunicativas entre indivíduos e instituições torna-se não apenas
essencial, mas imperativa, com o intuito de ampliar o interesse da sociedade na gestão
do território das áreas protegidas e assegurar a inclusão de todos os atores sociais nos
processos decisórios (Souza et al., 2022).

20
2.2 Percepções, atitudes e sentimentos relacionados as áreas protegidas
As distintas maneiras como as pessoas percebem e avaliam o meio ambiente são
sentidas individualmente ou no contexto de grupos sociais, e podem ser influenciadas
por diferentes experiências, aspirações e contextos socioeconômicos (McNeill, 2002).
Além disso, a percepção ambiental pode ser vista como um processo que vai além das
respostas sensoriais, incorporando impressões e sentimentos que refletem respostas
mentais

resultantes

experiências

individuais.

Essas

experiências

estão

intrinsecamente associadas a processos culturais, como destacado por Hoeffel e Fadini
(2007), e são enriquecidas pelas memórias e emoções que geram uma conexão
significativa das pessoas com os lugares (Tuan, 1980).
As áreas protegidas podem evocar tanto sentimentos positivos quanto negativos
nas pessoas. A interação com a natureza não apenas desencadeia emoções positivas,
contribuindo para a melhoria da saúde física, mental e psicológica (Velarde et al., 2007),
mas também atende a uma necessidade fundamental da sociedade. A percepção dos
atributos naturais dessas áreas pode significativamente influenciar a qualidade de vida
das pessoas (De Haan et al., 2014). Esses sentimentos positivos podem fomentar um
vínculo favorável com a natureza, promovendo um senso de pertencimento ao local
(Raffestin, 1993; Tuan, 1980) e estimulando comportamentos pró-conservação ambiental
(Hausmann et al., 2016).
Por outro lado, a insatisfação associada a experiências desfavoráveis na visitação
de áreas protegidas (Hausmann et al., 2020) ou questões relacionadas aos objetivos de
criação e gestão dessas áreas, como tamanho, restrições de uso do espaço e
regulamentações relacionadas à pesca, caça e proteção de espécies (Bragagnolo et al.,
2016; Souza et al., 2022), além da percepção de que as áreas protegidas prejudicam o
crescimento econômico, podem resultar na falta de apoio à sua criação, desinteresse e
alienação em relação à natureza (Hausmann et al., 2020).
É importante reconhecer que os sentimentos individuais não se limitam apenas a
expressões transitórias de emoções como alegria, raiva, interesse, tristeza e gratidão.
Em vez disso, os sentimentos das pessoas em relação as áreas protegidas podem causar

21
um impacto de longo prazo no comportamento das pessoas em relação ao meio ambiente
(Fredrickson, 2001).
A compreensão da percepção, conforme destacado por Lemberg (2010), engloba
a forma como as pessoas sentem, processam mentalmente e respondem às informações
derivadas do ambiente, sendo influenciada por características sociodemográficas,
atitudes e valores. Esses elementos têm o potencial de impactar diretamente a
experiência, a satisfação e os comportamentos associados às áreas protegidas (Rossi et
al., 2015). No contexto da conservação, a conexão entre os sentimentos das pessoas e
suas percepções desempenha um papel decisivo, uma vez que sentimentos adversos
podem fundamentalmente moldar a interpretação e interação das pessoas com o
ambiente natural. Um exemplo ilustrativo dessa dinâmica pôde ser observado por Maciel
(2015) no Parque Nacional da Tijuca, onde a concessão do Parque acarretou impactos
socioeconômicos negativos para a comunidade local, influenciando consequentemente
na percepção sobre a área protegida.
Associativamente, as percepções ambientais não são imparciais, refletindo, entre
outros aspectos, os interesses de grupos sociais distintos, atitudes, valores e visões de
mundo (Tuan, 1980). Pesquisas voltadas para compreender as relações homemnatureza, incluindo percepções, atitudes e comportamentos, são essenciais para uma
aplicação mais eficaz e adaptada das medidas de conservação em áreas protegidas
(Badola et al., 2012; Bragagnolo et al., 2016; Broad e Sanchirico, 2008; Souza et al.,
2022). Esses esforços de pesquisa estão sendo complementados por estudos de
percepções e opiniões sobre o meio ambiente nas redes sociais (Fink et al., 2020;
Hausmann et al., 2018), proporcionando uma compreensão mais abrangente das
dinâmicas contemporâneas nas relações homem-natureza.
No âmbito teórico, duas importantes bases dos estudos sobre percepções,
atitudes e comportamentos são a Teoria da Ação Racional (TAR) de Fishbein & Ajzen
(1975) e a Teoria do Comportamento Planejado (TCP) de Ajzen (1991). A TCP é uma
extensão da teoria da TAR. A TAR pressupõe que a intenção comportamental está
relacionada a dois fatores: (i) a atitude sobre o comportamento baseada nas crenças
comportamentais pessoais e (ii) normas subjetivas relacionadas às pressões sociais. No

22
entanto, a teoria recebeu muitas críticas por pressupor que a intenção comportamental
estava relacionada basicamente a fatores internos (julgamento pessoal favorável ou não
sobre determinado ato) e externos (como o ato será percebido pelo outro) (Moutinho e
Roazzi, 2010). Ajzen (1991) sugeriu então na TCP que além destes fatores, a intenção
comportamental era influenciada por um terceiro fator - o controle comportamental
presumido, onde se avalia os obstáculos ou a simplicidade em agir de determinada forma.
Essas teorias têm sido valiosas para compreender e prever as motivações por trás das
ações humanas relacionadas ao meio ambiente.
No contexto das redes sociais, onde as interações online são proeminentes, as
teorias comportamentais podem ser aplicadas de maneira pertinente. Considerando a
TAR e a TCP, é possível destacar que a atitude positiva ou negativa em relação a esses
comportamentos digitais é moldada por crenças pessoais de cada indivíduo. Além disso,
as expectativas sociais, que abrangem as normas subjetivas relacionadas às pressões
sociais, influenciam diretamente as decisões das pessoas sobre o que é aceitável
socialmente escrever, curtir e compartilhar. No entanto, para além dos fatores internos
(pessoais) e externos (sociais), a facilidade ou obstáculo percebido para realizar
determinados compartilhamentos de opiniões nas redes sociais desempenha um papel
crucial nas intenções e, por conseguinte, no comportamento real.
Dessa forma, ao compreender como as teorias comportamentais se aplicam ao
ambiente das redes sociais, podemos estabelecer uma ligação relevante com as
percepções sobre as APs. A forma como as pessoas interagem e compartilham opiniões
online pode influenciar não apenas seus comportamentos digitais, mas também se
estender para suas percepções e atitudes presenciais sobre questões ambientais,
incluindo as áreas protegidas.
2.3 Redes sociais e conservação ambiental
Tradicionalmente, as interações humanas com a natureza têm sido investigadas
por meio de pesquisas sociais. Recentemente, sugeriu-se que os enormes volumes de
dados gerados pelas pessoas nas redes sociais online e outras plataformas digitais
poderiam ser uma abordagem complementar utilizada para quantificar essas interações

23
em uma escala maior de população e acessibilidade geográfica (Di Minin et al., 2015) esse campo de estudo é chamado de culturômica da conservação (Ladle et al., 2016). A
culturômica refere-se à aplicação de métodos computacionais e análise de dados em
grande escala para estudar padrões culturais e mudanças ao longo do tempo, utilizando
grandes conjuntos de dados: textuais, vídeos, fotos e dados compartilhados em meio
digital (Ladle et al., 2017, 2016).
As redes sociais tornaram-se parte integrante do dia a dia moderno,
proporcionando oportunidades sem precedentes de comunicação, debate e partilha de
informação, para a compreensão das percepções e sentimentos do público (Sudhir and
Suresh, 2021). Uma das principais vantagens da análise de dados digitais é a grande
quantidade de informações geradas pelas pessoas sobre uma ampla gama de tópicos,
incluindo preferências políticas (Ceron et al., 2014), satisfação do cliente (Ahani et al.,
2019) e proteção natureza (Ladle et al., 2021; Di Minin et al., 2015; Souza et al., 2023).
Os dados das redes sociais podem ser recolhidos rapidamente, com custos financeiros
mais baixos do que outros métodos (por exemplo, questionários) (Becken et al., 2017) e
podem, portanto, ser usados como complementares – não invalidando o uso de
questionários e entrevistas face a face.
Em âmbito comercial, foi publicado um estudo sobre interesse e engajamento na
rede social Instagram que mostrou como as métricas de curtidas e comentários
influenciam na compra de produtos online (Aragão et al., 2016); outra pesquisa utilizando
as redes sociais como ferramenta investigativa mostrou por meio de análise de
sentimento com dados do Twitter, que existe uma capacidade notável dos meios de
comunicação social em prever os resultados eleitorais, bem como uma correlação
significativa entre os resultados das redes sociais e os resultados dos questionários
tradicionais (Ceron et al., 2014). Assim sendo, o crescente uso das redes sociais por um
público mais vasto de cidadãos vem aumentando fortemente a utilização dos dados da
internet como um dispositivo para investigar os interesses públicos e sentimentos da
sociedade com relação a natureza.
De fato, a utilização de dados online como do Google Trends e Wikipedia, e de
redes sociais como Twitter, Facebook e Instagram já demonstrou a sua aplicabilidade

24
para medir o interesse público e compreender as percepções dos utilizadores sobre
diversas questões de conservação (Almeida et al., 2022; Correia et al., 2021, 2017;
Guedes-Santos et al., 2021; Tenkanen et al., 2017). Um estudo específico destinado a
compreender o interesse público online em espécies ameaçadas analisou vídeos do
YouTube relacionados com o turismo de gorilas das montanhas (Gorilla beringei
beringei). O estudo de Otsuka e Yamakoshi (2020) mostra que vídeos que apresentam
contato físico ou interação com gorilas recebem mais visualizações e curtidas,
destacando a importância do monitoramento do interesse das pessoas pela natureza e
enfatizando a necessidade de melhorar a comunicação das ações voltadas à proteção
de espécies ameaçadas.
Em 2020, outro estudo analisou o interesse público da sociedade com relação aos
parques nacionais (PARNAs) no período da pandemia de COVID-19 em todo o mundo
(Souza et al., 2021). A pesquisa se concentrou, especificamente, nos impactos da
limitação da mobilidade das pessoas (através do isolamento social e de medidas de
lockdowns), considerando a forte relação causal do interesse público e a visitação nos
PARNAs; avaliando-os por meio do volume de pesquisas globais na Internet para cada
parque, utilizando a plataforma Google Trends. Os resultados revelaram que apesar da
redução global do interesse público nos PARNAs durante a pandemia, foi encontrada
uma diferenciação no comportamento tanto em nível local de cada parque como também
quando considerados em nível de país no qual está inserido (Figura 1). Esse
entendimento se mostra crucial não apenas para a gestão local, mas também para
iniciativas de marketing global, uma vez que as atividades dos visitantes podem resultar
em impactos diretos ou indiretos no meio ambiente (Toivonen et al., 2019). Esse
conhecimento pode ajudar a orientar os esforços de gestores e formuladores de políticas
para uma gestão mais eficaz das APs, identificando quais aspectos e eventos levam a
uma atitude pública negativa em relação às APs e abordando estratégias mais eficazes
para gerenciar tensões e promover mudanças em favor da conservação (Hausmann et
al., 2020; Hockings et al., 2006).

Figura 1: Boxplot representando a distribuição do volume de pesquisa relativa global para
parques nacionais na Austrália, Brasil, Espanha, Finlândia, Reino Unido, Índia, Tailândia, Estados
Unidos e África do Sul durante quatro períodos de tempo diferentes: janeiro a março de 20162019, março a julho de 2016-2020, janeiro a março de 2020 (antes da COVID-19) e março a julho
de 2020 (durante a Covid-19). Fonte: Souza et. al. (2021).

Além dos padrões de interesse, outra ferramenta interessante para analisar dados
digitais sobre interações entre humanos e natureza é a análise de sentimentos, que
aproveita os dados textuais de rede social para compreender os valores e as emoções
das pessoas em relação ao ambiente natural (Drijfhout et al., 2016). A análise de
sentimento emprega técnicas de processamento de linguagem natural (PLN) para
“analisar as opiniões das pessoas, sentimentos, avaliações, apreciações, atitudes e
emoções em relação a entidades como produtos, serviços, organizações, indivíduos,
questões, eventos, tópicos e seus atributos” (Liu, 2012, p.1), classificando o conteúdo de
dados textuais com base em percepções positivas, negativas ou neutras.
Essa abordagem tem uma ampla aplicação em marketing e outros campos de
estudo, e, mais recentemente, cientistas da conservação têm empregado essa
metodologia para avaliar os sentimentos das pessoas em relação à natureza, incluindo
gestão ambiental (Bhatt e Pickering, 2021), impactos do turismo na vida selvagem
(Otsuka et al., 2020) e as preferências, experiências e opiniões dos turistas (Hausmann
et al., 2020). Ao identificar emoções negativas, como frustração ou decepção, essa

26
abordagem pode indicar a necessidade de melhorias nas áreas de visitação de áreas
protegidas (Agyeman et al., 2019), bem como em práticas de gestão ou estratégias de
comunicação. Práticas polêmicas de gestão, como restrições ao uso do espaço e
medidas contra a caça ilegal (Lubbe et al., 2019), têm o potencial de gerar insatisfação,
falta de apoio e conflitos. Nesse sentido, a aplicação da análise de sentimento em dados
de rede social relacionados a áreas protegidas pode ser instrumental para identificar
problemas como os mencionados acima, contribuindo para a tomada de decisões
informadas e aprimoramento de estratégias de gestão (Agyeman et al., 2019).
2.3.1 Limitações das redes sociais como ferramenta investigativa
Mesmo considerando o potencial significativo das redes sociais para fornecer
percepções valiosas para a conservação, existem algumas limitações significativas
importantes de serem discutidas. Apesar de aproximadamente 90% dos brasileiros terem
acesso à Internet e ao menos 83,6% fazer uso de alguma rede social (IBGE, 2023), é
vital reconhecer que essa representação demográfica não abrange uniformemente todo
o país. No entanto, essa limitação não invalida o uso dessa abordagem investigativa, uma
vez que os resultados obtidos na pesquisa podem ter relevância e extrapolar o escopo
dos usuários de redes sociais para a sociedade (Ceron et al., 2014)
Outra limitação referente ao uso de dados online, envolve a presença significativa
de gírias, coloquialismos, acrônimos e emoticons em seus conteúdos textuais, o que
representa um desafio para a coleta e análise de dados. Além disso, a ausência da
geolocalização em dispositivos de alguns usuários pode limitar análises espaciais.
Portanto, todos os dados de rede social requerem limpeza extensiva e análise crítica
(Toivonen et al., 2019). Apesar destas limitações, os vastos volumes de dados gerados
por plataformas de redes sociais como o Twitter por exemplo, podem certamente fornecer
insights sobre as percepções de pessoas que não utilizam redes sociais (Ceron et al.,
2014) e, de forma crítica, podem avaliar padrões a nível macroscópico entre as interações
homem e natureza (Ladle, et al., 2021).
Partindo para a abordagem da análise de sentimento, apesar de seu grande
potencial para gerar percepções sobre a conservação da biodiversidade, uma limitação

27
significativa é a escassa disponibilidade de ferramentas e métodos para outros idiomas
além do inglês (Kaity and Balakrishnan, 2020). No contexto brasileiro, a lacuna de
estudos voltados a exploração dos sentimentos das pessoas sobre as áreas protegidas,
pode se dá pelos desafios metodológicos de aplicar essa ferramenta na língua
portuguesa. No entanto, quanto mais pesquisas nesta área forem desenvolvidas, mais
dados e resultados serão gerados a fim de fortalecer o uso desta importante ferramenta
metodológica.
Mesmo diante das limitações do uso de dados digitais para investigação, é de
conhecimento que os dados de culturomics são excepcionais em sua capacidade de
capturar, identificar e mapear sistematicamente interações homem-natureza em amplas
escalas espaciais e temporais. Diversos estudos já empregaram com sucesso dados de
redes sociais para informar comunicação científica (Papworth et al., 2015), explorar o
ecoturismo (Wibowo et al., 2019), avaliar o sentimento online em relação a espécies
ameaçadas (Fink et al., 2020) e compreender as percepções públicas sobre APs
(Hausmann et al., 2018; Souza et al., 2023). Entretanto, mais pesquisas são necessárias
para avaliar a relação específica entre o conteúdo das redes sociais e as APs, bem como
para investigar os sentimentos do público em relação a essas áreas. Tais pesquisas são
fundamentais para auxiliar na gestão da vasta e diversificada rede de APs brasileiras,
que têm sido cada vez mais atacados devido a uma tentativa de desmantelar as políticas
ambientais e de conservação na última década (Bernard et al., 2014; Fearnside, 2019;
Ferrante e Fearnside, 2019).
REFERÊNCIAS
Agyeman, Y.B., Aboagye, O.K., Ashie, E., 2019. Visitor satisfaction at Kakum National
Park in Ghana. Tourism Recreation Research 44, 178–189.
https://doi.org/10.1080/02508281.2019.1566048
Ahani, A., Nilashi, M., Yadegaridehkordi, E., Sanzogni, L., Tarik, A.R., Knox, K., Samad,
S., Ibrahim, O., 2019. Revealing customers’ satisfaction and preferences through online
review analysis: The case of Canary Islands hotels. Journal of Retailing and Consumer
Services 51, 331–343. https://doi.org/10.1016/j.jretconser.2019.06.014
Ajzen, I., 1991. The Theory of Planned Behavior.

28
Almeida, J.A.G.R., et al., 2022. Public awareness and engagement in relation to the
coastal oil spill in northeast Brazil. An Acad Bras Cienc 94, 1–10.
https://doi.org/10.37002/biobrasil.v12i2.2177
Aragão, F., Farias, F., Mota, M., Freitas, A., 2016. Curtiu, comentou, comprou. A mídia
social digital Instagram e o consumo. Revista Ciências Administrativas 22, 130–161.
https://doi.org/10.5020/2318-0722.2016.v22n1p130
Badola, R., Barthwal, S., Hussain, S.A., 2012. Attitudes of local communities towards
conservation of mangrove forests: A case study from the east coast of India. Estuar
Coast Shelf Sci 96, 188–196. https://doi.org/10.1016/j.ecss.2011.11.016
Becken, S., Stantic, B., Chen, J., Alaei, A.R., Connolly, R.M., 2017. Monitoring the
environment and human sentiment on the Great Barrier Reef: Assessing the potential of
collective sensing. J Environ Manage 203, 87–97.
https://doi.org/10.1016/j.jenvman.2017.07.007
Bernard, E., Penna, L.A.O., Araújo, E., 2014. Downgrading, downsizing, degazettement,
and reclassification of protected areas in Brazil. Conservation Biology 28, 939–950.
https://doi.org/10.1111/cobi.12298
Bhatt, P., Pickering, C.M., 2021. Public Perceptions about Nepalese National Parks: A
Global Twitter Discourse Analysis. Soc Nat Resour 34, 683–700.
https://doi.org/10.1080/08941920.2021.1876193
Bragagnolo, C., Gamarra, N.C., Malhado, A.C.M., Ladle, R.J., 2016. Proposta
Metodológica para Padronização dos Estudos de Atitudes em Comunidades Adjacentes
às Unidades de Conservação de Proteção Integral no Brasil. Biodiversidade Brasileira
6(1), 190–208.
BRASIL, 2000. Lei Federal Nº 9.985, de 18 de julho de 2000. Institui o Sistema Nacional
de Unidades de Conservação da Natureza e dá outras providências. Disponível em: <
https://www.planalto.gov.br/ccivil_03/leis/l9985.htm>. Acesso em: 10 de jan. de 2024.
BRASIL, 1998. Decreto legislativo Nº 2519, de 16 de março de 1998. Promulga a
Convenção sobre Diversidade Biológica, assinada no Rio de Janeiro, em 05 de junho
de 1992. Disponível em: < https://www.planalto.gov.br/ccivil_03/decreto/d2519.htm>.
Acesso em: 01 de fev. de 2024.
BRASIL, 1994. Decreto legislativo Nº 02, de 1994. Aprova o texto do Convenção sobre
Diversidade Biológica, assinada durante a Conferência das Nações Unidas sobre Meio
Ambiente e Desenvolvimento, realizada na Cidade do Rio de Janeiro, no período de 5 a
14 de junho de 1992. Disponível em: <
https://www2.camara.leg.br/legin/fed/decleg/1994/decretolegislativo-2-3-fevereiro-1994358280-publicacaooriginal-1-pl.html>. Acesso em: 01 de fev. de 2024.

29
Broad, K., Sanchirico, J.N., 2008. Local perspectives on marine reserve creation in the
Bahamas. Ocean Coast Manag 51, 763–771.
https://doi.org/10.1016/j.ocecoaman.2008.07.006
CBD. Convenção da Diversidade Biológica. 1992. Disponível em: <
https://www.planalto.gov.br/ccivil_03/decreto/1998/anexos/and2519-98.pdf>. Acesso
em: 01 de fev. de 2024.
Ceron, A., Curini, L., Iacus, S.M., Porro, G., 2014. Every tweet counts? How sentiment
analysis of social media can improve our knowledge of citizens’ political preferences
with an application to Italy and France. New Media Soc 16, 340–358.
https://doi.org/10.1177/1461444813480466
Correia RA, Ladle R, Jarić I, Malhado ACM, Mittermeier JC, Roll U, et al. Digital data
sources and methods for conservation culturomics. Conservation Biology. 2021;35:
398–411. doi:10.1111/cobi.13706
Correia, R.A., Jepson, P., Malhado, A.C.M., Ladle, R.J., 2017. Internet scientific name
frequency as an indicator of cultural salience of biodiversity. Ecol Indic 78, 549–555.
https://doi.org/10.1016/j.ecolind.2017.03.052
De Haan, F.J., Ferguson, B.C., Adamowicz, R.C., Johnstone, P., Brown, R.R., Wong,
T.H.F., 2014. The needs of society: A new understanding of transitions, sustainability
and liveability. Technol Forecast Soc Change 85, 121–132.
https://doi.org/10.1016/j.techfore.2013.09.005
Diegues, A.C. O mito moderno da natureza intocada / Antonio Carlos Santana Diegues.
— 3.a ed. —São Paulo : Hucitec Núcleo de Apoio à Pesquisa sobre Populações
Humanas e Áreas Úmidas Brasileiras, USP, 2000.
Di Minin, E. Di, Tenkanen, H., Toivonen, T., 2015. Prospects and challenges for social
media data in conservation science. Front Environ Sci 3, 1–6.
https://doi.org/10.3389/fenvs.2015.00063
Drijfhout, M., Kendal, D., Vohl, D., Green, P.T., 2016. Sentiment Analysis: ready for
conservation. Front Ecol Environ 14, 525–526. https://doi.org/10.1002/fee.1435
Engen, S., Hausner, V.H., Gurney, G.G., Broderstad, E.G., Keller, R., Lundberg, A.K.,
Murguzur, F.J.A., Salminen, E., Raymond, C.M., Falk-Andersson, J., Fauchald, P.,
2021. Blue justice: A survey for eliciting perceptions of environmental justice among
coastal planners’ and small-scale fishers in Northern-Norway. PLoS One 16, 1–20.
https://doi.org/10.1371/journal.pone.0251467
Fearnside, P.M., 2019. Setbacks under President Bolsonaro: A Challenge to
Sustainability in the Amazon. Sustentabilidade International Science Journal 1, 38–52.

30
Ferrante, L., Fearnside, P.M., 2019. Brazil’s new president and “ruralists” threaten
Amazonia’s environment, traditional peoples and the global climate. Environ Conserv.
https://doi.org/10.1017/S0376892919000213
Ferreira, J., Aragão, L.E.O.C., Barlow, J., Barreto, P., Berenguer, E., Bustamante, M.,
Gardner, T.A., Lees, A.C., Lima, A., Louzada, J., Pardini, R., Parry, L., Peres, C.A.,
Pompeu, P.S., Tabarelli, M., Zuanon, J., 2014. Brazil’s environmental leadership at risk:
Mining and dams threaten protected areas. Science (1979).
https://doi.org/10.1126/science.1260194
Fink, C., Hausmann, A., Di Minin, E., 2020. Online sentiment towards iconic species.
Biol Conserv 241. https://doi.org/10.1016/j.biocon.2019.108289
Fishbein, M. e Ajzen, I. 1975. Belief, attitude, intention, and behavior: An introduction to
theory and research. Addison-Wesley. 578p
Fredrickson, B.L., 2001. The role of positive emotions in positive psychology: The
broaden-and-build theory of positive emotions. American Psychologist 56, 218–226.
https://doi.org/10.1037/0003-066X.56.3.218
Gerhardinger, L.C., Godoy, E.A.S., Jones, P.J.S., Sales, G., Ferreira, B.P., 2011. Marine
protected dramas: The flaws of the Brazilian national system of marine protected areas.
Environ Manage 47, 630–643. https://doi.org/10.1007/s00267-010-9554-7
Guedes-Santos, J., Correia, R.A., Jepson, P., Ladle, R.J., 2021. Evaluating public
interest in protected areas using Wikipedia page views. J Nat Conserv 63.
https://doi.org/10.1016/j.jnc.2021.126040
Hausmann, A., Toivonen, T., Fink, C., Heikinheimo, V., Kulkarni, R., Tenkanen, H., Di
Minin, E., 2020. Understanding sentiment of national park visitors from social media
data. People and Nature pan3.10130. https://doi.org/10.1002/pan3.10130
Hausmann, A., Toivonen, T., Slotow, R., Tenkanen, H., Moilanen, A., Heikinheimo, V.,
Di Minin, E., 2018. Social Media Data Can Be Used to Understand Tourists’ Preferences
for Nature-Based Experiences in Protected Areas. Conserv Lett.
https://doi.org/10.1111/conl.12343
Hockings, M., Stolton, S., Leverington, F., 2006. Evaluating effectiveness : a framework
for assessing management effectiveness of protected areas, 2nd edition, Evaluating
effectiveness : a framework for assessing management effectiveness of protected areas,
2nd edition. https://doi.org/10.2305/iucn.ch.2006.pag.14.en
Hoeffel, J.L; Fadini, A.A.B. Percepção ambiental. In. Ferraro-Jr., L.A (org). Encontros e
Caminhos: Formação de Educadoras(es) Ambientais e Coletivos Educadores. Brasilia:
MMA, Departamento de Educação Ambiental, 2007. p. 253-263.
IBGE. Painel PNAD contínua. 2022. Disponível em: < https://painel.ibge.gov.br/pnadc/ >
Acesso em: 15 de janeiro de 2024.

31
Jepson, P.R., Caldecott, B., Schmitt, S.F., Carvalho, S.H.C., Correia, R.A., Gamarra, N.,
Bragagnolo, C., Malhado, A.C.M., Ladle, R.J., 2017. Protected area asset stewardship.
Biol Conserv 212, 183–190. https://doi.org/10.1016/j.biocon.2017.03.032
Kaity, M., Balakrishnan, V., 2020. Sentiment lexicons and non-English languages: a
survey. Knowl Inf Syst 62, 4445–4480. https://doi.org/10.1007/s10115-020-01497-6
Ladle, Richard J.; Souza, Carolina N.; Correia, R., 2021. Culturomics for (not against!)
protected areas In. Biol Conserv 256, 109197.
https://doi.org/10.1016/j.biocon.2021.109015
Ladle, R.J., Correia, R.A., Do, Y., Joo, G.J., Malhado, A.C.M., Proulx, R., Roberge, J.M.,
Jepson, P., 2016. Conservation culturomics. Front Ecol Environ 14, 269–275.
https://doi.org/10.1002/fee.1260
Ladle, R.J., Jepson, P., Correia, R.A., Malhado, A.C.M., 2017. The power and the
promise of culturomics. Front Ecol Environ 15, 290–291.
https://doi.org/10.1002/fee.1506
Lemberg D., “Environmental Perception” in Warf, B., Encyclopedia of Geography, Sage,
2010.
Liu, B., 2012. Sentiment analysis and opinion mining. Morgan & Claypool.
Lubbe, B.A., du Preez, E.A., Douglas, A., Fairer-Wessels, F., 2019. The impact of rhino
poaching on tourist experiences and future visitation to National Parks in South Africa.
Current Issues in Tourism. https://doi.org/10.1080/13683500.2017.1343807
Macedo, H.S., Medeiros, R.P., 2018. Rethinking governance in a Brazilian multiple-use
marine protected area. Mar Policy 0–1. https://doi.org/10.1016/j.marpol.2018.08.019
Maciel, G.G., 2015. Mercantilização da cidade do Rio de Janeiro e suas implicações na
gestão de unidades de onservcação: um estudo sobre a concessão do Setor
Paineras/Corcovado (Parque Nacional da Tijuca - RJ) e os efeitos sobre os moradores
das favelas do Cerro Corá e do Guararapes. Pontifícia Universidade Católica do Rio de
Janeiro, Rio de Janeiro.
Maharani Wibowo, J., Dra Sri Muljaningsih, S., Dias Satria, Ms., Wibowo, J.M.,
Muljaningsih, S., Satria, D., 2019. Tripadvisor sentiment analysis: the policy of
ecotourism competitiveness from bromo, tengger, and semeru national park.
International Journal of Business, Economics and Law 20, 18–24.
Maretti, Cláudio C; Catapan, Marisete; Abreu, Maria Jasylene;Oliveira, J.E.Dantas.,
2012. Áreas protegidas: definições, tipos e conjuntos – reflexões conceituais e diretrizes
para a gestão.

32
McNeill, J., 2002. Something New Under the Sun: An Environmental History of the
Twentieth-Century World. Oxford University Press 36, 183–185.
https://doi.org/10.1353/jsh.2002.0109
MMA, 2023. Painel de Unidades de Conservação. Disponível em:
<https://cnuc.mma.gov.br/powerbi> Acesso em: 06 de novembro de 2023.
Moutinho, K., Roazzi, A., 2010. As teorias da ação racional e da ação planejada:
relações entre intenções e comportamentos. Aval. psicol 9, 279–287.
Otsuka, R., Yamakoshi, G., Id, R.O., Yamakoshi, G., 2020. Analyzing the popularity of
YouTube videos that violate mountain gorilla tourism regulations. PLoS One 15, 1–20.
https://doi.org/10.1371/journal.pone.0232085
Papworth, S.K., Nghiem, T.P.L., Chimalakonda, D., Posa, M.R.C., Wijedasa, L.S.,
Bickford, D., Carrasco, L.R., 2015. Quantifying the role of online news in linking
conservation research to Facebook and Twitter. Conservation Biology 29, 825–833.
https://doi.org/10.1111/cobi.12455
Raffestin, C., 1993. O que é Território?, in: Por Uma Geografia Do Poder. pp. 143–163.
Rossi, S.D., Byrne, J.A., Pickering, C.M., Reser, J., 2015. “Seeing red” in national parks:
How visitors’ values affect perceptions and park experiences. Geoforum 66, 41–52.
https://doi.org/10.1016/j.geoforum.2015.09.009
Rylands, A.B., Brandon, K., 2005. Brazilian protected areas. Conservation Biology.
https://doi.org/10.1111/j.1523-1739.2005.00711.x
Selva, V. S. F.; Souza, C. N.; Gouveia, R. L.; Santos, E. C. S. Práticas turísticas em
áreas protegidas: um olhar sobre a Área de Proteção Ambiental - APA Costa dos
Corais, Brasil. In: Giovanni Seabra. (Org.). Terra - paisagens, solos, biodiversidade e
os desafios para um bom viver. 1ed.Ituiutaba: Barlavento, 2016, v. 1, p. 1024-1034
Silva, J.M.C. da, Dias, T.C.A. de C., Cunha, A.C. da, Cunha, H.F.A., 2021. Funding
deficits of protected areas in Brazil. Land use policy 100.
https://doi.org/10.1016/j.landusepol.2020.104926
Souza, C.N., Almeida, J.A.G.R., Correia, R.A., Ladle, R.J., Carvalho, A.R., Malhado,
A.C.M., 2023. Assessing Brazilian protected areas through social media: Insights from
10 years of public interest and engagement. PLoS One 18.
https://doi.org/10.1371/journal.pone.0293581
Souza, C.N., de Barros, E.L.S.F.C., Dantas, I.F. V., Bragagnolo, C., Malhado, A.C.M.,
Selva, V.F., 2022. Inclusion and governance in the managing Council of the Costa dos
Corais Environmental Protection Area. Ambiente e Sociedade 25.
https://doi.org/10.1590/1809-4422ASOC20210074R1VU2022L3AO

33
Souza, C.N., Rodrigues, A.C., Correia, R.A., Normande, I.C., Costa, H.C.M., GuedesSantos, J., Malhado, A.C.M., Carvalho, A.R., Ladle, R.J., 2021. No visit, no interest:
How COVID-19 has affected public interest in world’s national parks. Biol Conserv 256.
https://doi.org/10.1016/j.biocon.2021.109015
Sudhir, P., Suresh, V.D., 2021. Comparative study of various approaches, applications
and classifiers for sentiment analysis. Global Transitions Proceedings 2, 205–211.
https://doi.org/10.1016/j.gltp.2021.08.004
Tenkanen, H., Di Minin, E., Heikinheimo, V., Hausmann, A., Herbst, M., Kajala, L.,
Toivonen, T., 2017. Instagram, Flickr, or Twitter: Assessing the usability of social media
data for visitor monitoring in protected areas. Sci Rep 7. https://doi.org/10.1038/s41598017-18007-4
Toivonen, T., Heikinheimo, V., Fink, C., Hausmann, A., Hiippala, T., Järv, O., Tenkanen,
H., Di Minin, E., 2019. Social media data for conservation science: A methodological
overview. Biol Conserv 233, 298–315. https://doi.org/10.1016/j.biocon.2019.01.023
Tuan. Y. Topofilia: um estudo da percepção, atitudes e valores do meio ambiente.
Trad.: Lívia de Oliveira. Londrina: Eduel, 2012.342p.1980-2012.
Velarde, M.D., Fry, G., Tveit, M., 2007. Health effects of viewing landscapes Landscape types in environmental psychology. Urban For Urban Green 6, 199–212.
https://doi.org/10.1016/j.ufug.2007.07.001
Watson, J.E.M., Dudley, N., Segan, D.B., Hockings, M., 2014. The performance and
potential of protected areas. Nature 515, 67–73. https://doi.org/10.1038/nature13947

34
3 OBJETIVOS
3.1 Objetivo geral

Compreender o interesse e percepções em relação às áreas protegidas brasileiras por
meio das redes sociais e da análise de sentimentos.

3.2 Objetivos específicos

Quantificar o volume de postagens no Twitter relacionadas às áreas protegidas
brasileiras;

ii.

Mapear geograficamente a origem das postagens realizadas no Twitter sobre as
áreas protegidas brasileiras;

iii.

Analisar e identificar quais áreas protegidas brasileiras geram maior volume de
publicações e engajamento no Twitter;

iv.

Investigar a relação entre o número de publicações e o engajamento dos
usuários que postam sobre áreas protegidas no Twitter;

Identificar os tópicos mais discutidos nas postagens sobre áreas protegidas
brasileiras no Twitter;

vi.

Classificar os sentimentos públicos (positivos, neutros e negativos) expressos no
conteúdo textual do Twitter sobre os parques nacionais brasileiros;

vii.

Identificar os principais tópicos que contribuem para as percepções negativas
relacionadas aos parques nacionais brasileiros no Twitter.

35
4 ASSESSING BRAZILIAN PROTECTED AREAS THROUGH SOCIAL MEDIA:
INSIGHTS FROM 10 YEARS OF PUBLIC INTEREST AND ENGAGEMENT 3

Carolina Neves Souza1*, João A. G. R. Almeida2, Ricardo A. Correia3,4,5, Richard J.
Ladle1,6,#a, Adriana R. Carvalho7, Ana C. M. Malhado1

1 Programa de pós-graduação em Diversidade Biológica e Conservação nos Trópicos.

Instituto de Ciências Biológicas e da Saúde, Universidade Federal de Alagoas, Maceió,
Alagoas, Brasil
2 Instituto de Computação, Universidade Federal de Alagoas, Maceió, Alagoas, Brasil
3

Helsinki Lab of Interdisciplinary Conservation Science (HELICS), Department of

Geosciences and Geography, University of Helsinki, Helsinki, Finland
4 Helsinki Institute of Sustainability Science (HELSUS), University of Helsinki, Helsinki,

Finland
5 Biodiversity Unit, University of Turku, Turku, Finland
6 Centro de Investigação em Biodiversidade e Recursos Genéticos (CIBIO), Universidade

do Porto, Vairão, Portugal
#a Current Address:

BIOPOLIS Program in Genomics, Biodiversity and Land Planning,

CIBIO, Vairão, Portugal
7 Departamento de Ecologia, Universidade Federal do Rio Grande do Norte, Natal, Rio

Grande do Norte, Brasil

3 Artigo publicado na revista PLOS ONE. Souza, C.N., Almeida, J.A.G.R, Correia R.A, Ladle R.J,

Carvalho A.R, Malhado A.C.M. 2023. Assessing Brazilian protected areas through social media: Insights
from 10 years of public interest and engagement. PLoS ONE 18(10): e0293581.
https://doi.org/10.1371/journal.pone.0293581

36
4.1 Abstract
Social media platforms are a valuable source of data for investigating cultural and
political trends related to public interest in nature and conservation. Here, we use the
micro-blogging social network Twitter to explore trends in public interest in Brazilian
protected areas (PAs). We identified ~400,000 Portuguese language tweets pertaining to
all categories of Brazilian PAs over a ten-year period (1 January 2011 - 31 December
2020). We analysed the content of these tweets and calculated metrics of user
engagement (likes and retweets) to uncover patterns and drivers of public interest in
Brazilian PAs. Our results indicate that users / tweets mentioning PAs remained stable
throughout the sample period. However, engagement with tweets grew steeply,
particularly from 2018 onward and coinciding with a change in the Brazilian federal
government. Furthermore, public interest was not evenly distributed across PAs; while
national parks were the subject of the most tweets, mainly related to tourism activities,
tweets related to conflicts among park users and managers were more likely to engage
Twitter users. Our study highlights that automatic or semi-automatic monitoring of social
media content and engagement has great potential as an early warning system to identify
emerging conflicts and to generate data and metrics to support PA policy, governance
and management.
4.2 Introduction
Protected areas (PAs) are a key tool for biodiversity conservation [1]. In Brazil,
these areas are not only responsible for protecting different ecosystems, habitats and
endangered species, they also safeguard important cultural and socioeconomic values
[2]. To align social and economic aims, in addition to conservation and recreation, the
Brazilian system includes two broad categories of PAs: strictly protected and sustainable
use. Indigenous lands also provide an important contribution to biodiversity conservation,
although they are not formally recognized as PAs in the National Protected Areas System
(SNUC) [3] and are regulated by different legislation [4]. The Brazilian PA system is one
of the largest in the world, but is facing a range of challenges that threaten its integrity and
long-term sustainability. Among the most pressing of these challenges are: i) its reliance

37
on outdated top-down governance systems that do not sufficiently allow for the direct
participation of society [5]; ii) the long term and persistent institutional crisis facing the
federal biodiversity conservation agency [6]; iii) a lack of transparency in management
actions and in communicating the importance of these areas, which tends to result in
increasing environmental crimes and corresponding diminishment of management
effectiveness and monitoring of these areas [7], and; iv) a growing funding deficit, which
weakens PAs that are not able to cover their management costs [8]. In summary, fostering
a stronger connection between Brazilian PAs and society is crucial to avoid them being
perceived as opportunity costs by citizens and politicians [9]. In addition, understanding
how people interact with these PAs can provide important insights on how to increase
society's support for conservation efforts.
Individuals interact with PAs in a wide range of ways, generating diverse values
and evoking different interests and feelings [10]. In Brazil, depending on the PA category,
citizens and visitors can engage in a wide range of activities including recreation,
research, developing environmental education activities, or simply visiting and enjoying
the iconic landscapes and biological spectacles. Such interactions have demonstrable
psychological and physical benefits to humans and can promote well-being [11].
Nevertheless, Brazilian PAs are primarily configured for environmental conservation, with
varying use restrictions that can contribute to a lack of interest from wider society and a
general alienation from nature [12]. Given that public support is critical for the legitimacy
of PAs [7], understanding human interactions, sentiment and public interest in PAs is
essential for developing effective strategies to attract societal support and for supporting
decision-makers and researchers in conservation planning, financing and public
communication activities [13–15].
Traditionally, human interactions with nature have been investigated through social
surveys which are necessarily costly and limited in scale. It has recently been suggested
that the huge volumes of data generated by social media and other digital platforms could
be a complementary approach utilised to quantify these interactions at a larger scale of
population and geographic accessibility [16] - this field of study is called conservation
culturomics [14]. In comparison to questionnaires, social media analysis has the potential
to generate large amounts of data, at a lower financial cost, and on a larger geographical

38
scale [13]. This in no way invalidates the continued use of questionnaires as a valuable
methodological tool; data generated from analysis of social networks have many intrinsic
biases [17] and, critically, does not capture the attitudes and behaviours of
communities/individuals with limited access to the internet and/or those that do not use
social networks [12,18]. Another concern is the reliability of how the data is recorded and
made available by the system. According to [19], inconsistent data measurements by the
system can also undermine internal validity, making it difficult to infer causality from the
responses. The authors therefore suggest that researchers familiarise themselves with
the system used in the survey in order to validate the results. Although about 90% of
Brazilians have access to the Internet, this demographic representation does not cover
the entire country. Nevertheless, this limitation does not invalidate the usefulness of this
investigative approach, since the results obtained in the research may often have
relevance beyond the scope of social network users [18]. Culturomic data is unrivalled in
its potential for systematically capturing, identifying and mapping human-nature
interactions at large spatial and temporal scales [17]. Indeed, social media data has
already been used successfully to inform science communication [20,21], investigate
ecotourism in high conflict environments [22], assess online sentiment towards threatened
species [23], enhance public awareness of wildlife conservation [24], and to better
understand public perceptions and feelings related to PAs [12,13,25]. In Brazil, Google
Trends data has previously been used to assess public interest and internet salience in
relation to Brazilian PAs [26]. However, more studies are needed to assess the
relationship between social media content and PAs, as well as to investigate public
sentiment towards these areas.
Here, we use data from the social media platform Twitter (recently renamed X) to
investigate public attitudes and interest in Brazilian PAs. Twitter is one of the most popular
social media and microblogging platforms with over 436 million active users worldwide in
2021 [27]. In Brazil, Twitter has up to 14.1 million active users, who posted millions of
comments (so-called “tweets”) every day containing thoughts and opinions of up to 280
characters [28]. Despite the character limit, the inclusion of links significantly enhances
the informational content and potential impact of tweets, as it enables individuals to access
additional resources, broaden their knowledge, and engage in more enriching and

39
informed discussions. The company has promoted itself as the right place to learn more
about “what's going on” and “what people are talking about right now”. Twitter is heavily
used by journalists, scientists, politicians, managers, and wider society [29,30] to spread
information, promote public discourse, and thus serves as a potentially sensitive
barometer of public opinion [31]. However, it is important to note that due to its open nature
Twitter can also be responsible for the dissemination of misinformation and the spread of
fake news.
In this study we analysed ten years (2011-2020) of public tweets in Portuguese
language that contained content related to Brazilian PAs with the objective of answering
the following questions: (i) what is the volume of posts on Twitter about Brazilian PAs? (ii)
where are people communicating about Brazilian PAs? (iii) which types of Brazilian PAs
generate more posts and engagement (e.g., other users reacting to the original posts
through ‘likes’ and ‘retweets’)? (iv) what is the relationship between the number of posts
and the engagement of users who post about PAs? (v) what are the most discussed topics
in the posts? Answering these questions through digital data analytics can inform targeted
communication strategies, improve public engagement and foster a deeper understanding
of biodiversity conservation, leading to more effective and impactful conservation
initiatives.
4.3 Material and Methods

4.3.1 Brazilian protected areas

Brazil has enormous biodiversity and an extensive system of conservation units,
protected spaces that are part of the Brazilian territory and that are managed to conserve
its ecological, historical, geological, and cultural heritage [32]. Since Brazil committed to
international programmes such as the Convention on Biological Diversity (CBD) and
national targets aimed at conservation, the country’s PAs system has rapidly expanded.
Although the system of PAs in Brazil was not directly created in response to the CBD, the
Convention served as a backdrop for the establishment of the National System of PAs
(SNUC). (This occurred because the initial versions of the law that instituted the SNUC

40
predated the CBD). During its passage through the National Congress, which took
approximately 12 years, some of the guidelines from the Convention were incorporated
into the text of this legal framework, thereby making it the primary instrument focused on
biodiversity conservation in Brazil [33]. The consolidation of the various norms regarding
Conservation Units in Brazil was not straightforward due to frequent disagreements
between conservationist and preservationist perceptions of these areas [34]. However,
after a long process of discussions among technicians, researchers, and public bodies,
the National System of Conservation Units (Sistema Nacional de Unidades de
Conservação, commonly referred to by its acronym SNUC) was unified under a single law
(see [3]).
The SNUC recognizes 12 categories of conservation units, separated according to
their management objectives and types of use. These categories fall into two major
groups: strict protection and sustainable use PAs [3]. Most, but not all, of Brazil’s PAs are
documented in the National Registry of Protected Areas (Cadastro Nacional de Unidades
de Conservação - CNUC). The CNUC currently includes 2,659 PAs, including marine and
private PAs, managed at federal, state, and municipal levels. These areas cover about
18.80% of the continental area and 26.48% of the marine area of Brazil [35]. Between
2003 and 2009, Brazil was solely responsible for 74% of the global increase in PA
coverage (km2), mainly due to several large PAs created in Amazonia [36]. Despite
Brazil’s leading role in global conservation and the immense success of its PA
programme, Brazilian politicians and decision-makers seem to be increasingly viewing
PAs as opportunity costs that limit economic development, leaving many PAs vulnerable
to downgrading, downsizing or degazettement (PADDD) [37]. In this context,
demonstrating public support for PAs and revealing their true value to society is an
essential step to ensure their long-term sustainability [9].
4.3.2 Data collection
Digital data for conservation culturomics analysis can be collected from different
sources (e.g., texts, videos, images, songs) [38]. Based on the framework suggested by
[17], we analysed the content of, engagement with and author characteristics of publicly
available Twitter posts about Brazilian PAs in Portuguese language. The data mining

41
techniques involved data collection, cleaning, processing and analysis (see Fig 1). In this
study, we utilised the Twitter v2 API to collect all the data. Twitter's Academic access to
its v2 API has provided us with the opportunity to gather up to 10 million tweets per month,
which is a significant increase of 20 times compared to what was previously possible with
the standard v1.1 API [39]. Moreover, this access allows us to retrieve older conversation
histories, making Twitter a rich and accessible source of data for textual content analysis.
As such, it becomes a valuable tool for gaining insights and a better understanding of
public discussions and perceptions about Brazilian PAs. The code for data mining was
developed using the Python language program v.3.9 (http://www.python.org) and was
based on the Full-Archive-Search API node example from Twitter's official repository.

Fig 1. Methodological flowchart. Methodological flowchart from data collection to results. The
flowchart shows all the steps used during the research: data collection, cleaning, processing, and
analysis.

The second step was to define which content we would like to collect from Twitter.
A query of 18 keywords was defined in the Portuguese languaged to extract the tweets
(not the retweets). : 1- Parque nacional (National park); 2- Parque Estadual (State park);
3- Parque natural municipal (Natural municipal park); 4- Parque municipal (Municipal
park); 5- Estação Ecológica (Ecological station); 6- Reserva Biológica (Biological reserve);
7- Monumento Natural (Natural monument); 8- Refúgio da Vida Silvestre (Wildlife refuge);
9- Reserva Extrativista (Extractivism reserve); 10- Área de proteção ambiental
(Environmental protected area); 11- Floresta nacional (National forest); 12- Floresta
estadual (State forest); 13- Floresta municipal (Municipal forest); 14- Reserva de
desenvolvimento sustentável (Sustainable development reserve); 15- Área de interesse
relevante (Area of relevant interest); 16- Reserva particular do patrimônio natural (Private
natural reserve); 17- Unidade de conservação (Conservation unit); 18- Área protegida
(Protected area).
All 12 categories of Brazilian PAs were included in the query. PAs with
management levels in their name were also added to the query, for example, “national
park”, “state park”, and “municipal park”. Furthermore, to collect the tweets that do not
mention the names of PAs in their textual content, the query also included the two
keywords: conservation unit and protected area (unidade de conservação e área
protegida). We also considered adding to the set of terms used in the query the acronyms
referring to each PA category, for example APA for Área de Proteção Ambiental
(Environmental Protection Area). However, a preliminary analysis of data collected using
the acronyms revealed these mostly relate to other topics such as slang, celebrity names
and other words that did not result in tweets discussing PAs. Based on this, even though
it resulted in reducing our sample size, we decided not to include the acronyms in the
query to reduce bias and noise.
The collection of the textual content of retweets was also avoided to prevent
possible biases in the results of the multiple counting, and in the salience of topics, since
they generally represent repetitive copies of original tweets. Instead of collecting the
textual content of retweets, we collected quantitative data from the original comments that
contained the number of times the message was retweeted, received likes or comments.

43
In addition, the use of keywords in our query made it possible to identify and collect all
messages, including retweets that had comments related to the original tweet, that
mentioned any Brazilian PAs. The tweets were retrieved from the API between April 15th
and April 30th, 2021. The sampling period encompassed tweets posted from January 1st,
2011, to December 31st, 2020. In total, before the data cleaning and filtering process,
421,254 tweets were collected.
The following information was collected from each tweet: (i) author (name,
username, and whether it is verified); (ii) date of publication; (iii) geographical data
(latitude, longitude, city, country); (iv) publication data (number of likes, retweets, replies
and whether it is a reply to another user); and (v) the text of the tweet. The tweets were
downloaded in JSON format, by year, with a maximum of 500 tweets per page (as per the
limit set by the Twitter API). After converting the JSON pages to a single CSV file, the
data cleaning process was performed. Initially, a filter was applied to the geographical
metadata provided by Twitter (country and country_code columns) to identify tweets from
countries other than Brazil. Subsequently, a manual verification of each foreign tweet was
conducted to eliminate those that did not correspond to Brazilian PAs. Our final validated
list contained a total of 402,508 tweets about Brazilian PAs (see all IDs collected from
Twitter in https://github.com/jagra26/Brazilian-PAs-on-twitter). Tweets usually contain
URLs, emojis, and emoticons, so the dataset needs to be cleaned before analysis. Text
cleaning was performed using the R language program (R Core Team, 2017), with the tm
package [40]. Specifically, we used the function ‘tm_map’ to convert all text to lower case
and remove any hashtags, URLs, symbols (like 🡺,■,◆), numbers, and Portuguese stop
words present in the text. In terms of data collection, it's worth noting that the availability
of the APIs for academic usage has been volatile recently, as the platform is now referred
to as X and has undergone a policy change in data access which has restricted access.
4.3.3 Data analysis
We aimed to understand the public interest in Brazilian PAs based on the number
of tweets and public engagement with those posts. First, we compared the volume of
tweets about Brazilian PAs on a temporal scale of 2011 to 2020 with the number of users

44
who post about PAs. To do this, we counted the number of tweets, number of unique
users, and the total number of ‘likes’ and retweets by year to create a line graph. To
explore the relationship between the volume of tweets and engagement, we summarised
the number of tweets and the average number of ‘likes’ and retweets per post at the user
level. We calculated the bootstrapped mean engagement over time using 1000 samples
and a confidence interval of 95% with function ‘smean.cl.boot’ from R package Hmisc [41],
and generated a line plot depicting the disparity between user types (verified and not
verified). We also bootstrapped the Spearman’s correlation between number of tweets
and mean engagement per tweet using 1000 samples and 95% confidence intervals with
function ‘spearmanRho’ from R package rcompanion [42].
We also mapped the number of Tweets about Brazilian PAs based on location
information provided by Twitter’s users. To identify the geolocation of tweets related to
Brazilian PAs, we employed a keyword-based (see query used in the data collection
section) sampling strategy that increased the likelihood of capturing tweets about the
targeted PAs originating from various locations [25]. The metadata of the tweets provided
geographic data such as coordinates, countries, and cities. However, the consistency of
this information was limited, with many tweets containing only a single geographic
coordinate (e.g., the absence of country-level data despite the availability of coordinates).
To overcome this limitation, we adopted a similar approach to [43], where we utilised the
OpenCage Geocoding API search engine (https://opencagedata.com/) for reverse
geocoding. This process allowed us to obtain comprehensive state, and country
information, based on the coordinate metadata provided by Twitter. Ultimately, our
geospatial dataset consisted of 62,924 tweets (15.63%) with location data (geographic
coordinates and states and countries data). Using the folium library [44] in Python version
3.9 (http://www.python.org/), we created a choropleth map where the colour intensity of
each territory corresponded to the number of tweets posted.
It was necessary to implement a filtering process to assess which Brazilian PAs
had the most posts, as the majority of tweets did not explicitly mention the full names of
the PAs in their text content. To accomplish this, we filtered our dataset using the proper
names of the PAs from the National Register of Protected Areas (MMA/CNUC 2021),
without considering the specific category of the PA in the search, using the VLOOKUP

45
function through Microsoft Excel software (Microsoft 365 2020): =VLOOKUP
(search_value, SEARCH (search_text, no_text, [start_num]), search_text). Exceptions to
the filtering process using the full name with the category type were made for those areas
that had the same name with different categories, such as "Tamoios Ecological Station''
and "Tamoios Environmental Protection Area." The list of names used for filtering can be
accessed on the repository (https://github.com/jagra26/Brazilian-PAs-on-twitter). The final
dataset used in this analysis has 189,294 tweets containing PAs in CNUC. Using the new
dataset, we generated a lollipop chart using summarised data on the number of posts and
average engagement per PA to compare the differences between the public interest of
each area.
Finally, to group tweets based on similarity in content and to identify the main topics
Twitter users discuss concerning Brazilian PAs, we applied the agglomerative clustering
technique, known as AGNES (Agglomerative Nesting). The clustering method used to
calculate the distances was the "complete" method, based on the word composition of
posts in the dataset [45]. The complete method uses the “largest dissimilarity between a
point in the first cluster and a point in the second cluster (furthest neighbour method)” [46].
According to the study conducted by [47], in order to ensure that the final clusters became
neither excessively broad (involving few clusters covering PAs that share the most
common words) nor excessively specific (resulting in numerous clusters addressing PAs
that share less frequent words), the authors carried out several simulations with a range
of 2 to 10 clusters (k).
During these simulations, the most frequent words in each cluster were examined
in each scenario in order to identify the most prominent characteristics of each cluster in
relation to the Brazilian PAs. The cluster analysis produced a dendrogram in which the
distances between the branches reflect the similarity between each tweet. This procedure
made it possible to define the theme of each cluster (5) based on the similarity of the
words grouped in the dendrogram.With exception of the choropleth map, all analyses
were performed in the R programming language. The following packages were used in
the analysis: dplyr package was used for the data manipulation [48]; to create the graphics
we used the packages ggplot2 [49] and plotly [50]; and, finally, for the hierarchical
clustering analysis, we used the packages cluster [45] and factoextra [51]. All R and

46
Python scripts used for data collection and analysis are available on our repository page
(https://github.com/jagra26/Brazilian-PAs-on-twitter).
4.4. Results

4.4.1 Volume of tweets about Brazilian PAs

We mined a total of 402,508 valid tweets posts about Brazilian PAs between 2011
and 2020, which were those that successfully passed through the data cleaning process.
The number of posts and the number of users tweeting about the topic was relatively
similar throughout the decade analysed, with only minor decreases in the number of posts
after 2016 (Fig 2).

Fig 2. Volume of posts, users and engagement of Brazilian protected areas. Line graph
representing the volume of users, tweets, and public engagement (likes + retweets) about
Brazilian protected areas during the period from 01 January 2011 to 31 December 2020.

The average number of posts and users posting about PAs in Brazil was around
40,000 posts and around 17,000 users over the 10 years analysed. Metrics of
engagement (likes + retweets) with posts about PAs started increasing in 2016, but grew

47
steeply after 2018. Estimates of the mean engagement per tweet increased from 0.182
(bootstraped 95% C.I.: 0.165-0.199) in 2011 to 15.3 (bootstrapped 95% C.I.: 11.9-19.6)
(Fig 3).

Fig 3. Bootstrapped estimates of mean post engagement per year. Figure shows the results
of a Bootstrap analysis representing the mean engagement (likes and retweets) with posts about
Brazilian protected areas published between 2011 and 2020.

Compared to 2018, the number of people actively posting about PAs in Brazil in
2019 increased by 40%, posts about what is happening in these areas increased by 31%,
and public participation increased by 255% (Fig 4).

Fig 4. Public interest related to the Brazilian protected areas over the years. Boxplot
represents the distribution of the volume of likes for Brazilian protected areas during the period
from 01 January 2011 to 31 December 2020. The lower and upper box limits correspond to the
first and third quartiles (the 25th and 75th percentiles); a black bar inside the box indicates the
median. The volume of tweets was log-transformed (natural logarithm).

4.4.2 Users characteristics

We identified 130,742 Twitter users who have posted content about Brazilian PAs.
Some of these Twitter users have many followers, including news media, politicians,
celebrities, travel agencies, and a range of international people and organisations such
as WWF and UNESCO. These are frequently classified by Twitter as ‘users of public
interest’ and, in most cases, received a verification seal. We explored the relationship
between user type (verified or not) and the number of tweets about Brazilian PAs and the
public interest in those posts (Fig 5). In general, users who post more about Brazilian PAs
receive more engagement (likes + retweets). Indeed, there was a small but positive
correlation (Spearman’s rho = 0.107; bootstrapped 95% C.I.: 0.102-0.112) between the
number of tweets per user and mean engagement per tweet. However, it is worth noting
that when we compared engagement between verified and not verified users, our

49
bootstrap analysis showed that the estimates of the average engagement per tweet of an
not verified user increased from 0.354 (bootstrapped 95% CI: 0. 305-0.409) in 2011 to
22.3 (bootstrapped 95% CI: 12.9-34.5) in 2020, while the average engagement per tweet
of verified users increased from 7.00 (bootstrapped 95% CI: 4.48-10.1) in 2011 to 651.8
(bootstrapped 95% CI: 388.6-1003.9) (Fig. 5). For example, of the 10 users who have
posts with an average engagement of more than 5,000 likes and retweets, seven are
verified profiles and have a considerable number of followers (>15,000 followers).
Interestingly, the most of these users, despite their large following, posted only one or two
tweets related to Brazilian PAs during the analysed ten-year period.

Fig 5. Bootstrapped estimates of mean post engagement per year and per user type. Figure
shows the results of a Bootstrap analysis representing the mean engagement (likes and retweets)
with posts about Brazilian protected areas, per type of user (verified and not verified) published
between 2011 and 2020.

4.4.3 Geographic focus of tweets

The large majority of tweets (provided by the Twitter coordinates metadata) about
Brazilian PAs originate in Brazil, as expected, but we also found mentions from countries
bordering Brazil, Costa Rica, Portugal, and the United States (Fig 6). This result is similar
to

the

spatial

distribution

found

lusophone

and

associated

countries

(https://www.cplp.org/). In Brazil, the state of Rio de Janeiro received the highest number
of georeferenced tweets about PAs (12,585), followed by Minas Gerais (11,339), São

50
Paulo (8,837), and Paraná (5,434). Nevertheless, when we examine which PAs were most
georeferenced by Twitter users, we can see that out of the top 100 geo-referenced areas,
64% are urban parks. The most geotagged National Park was Iguaçu National Park
(Paraná state) with 2,747 tweets, followed by Tijuca National Park (Rio de Janeiro state)
with 1,300 tweets and Serra dos Órgãos National Park (Rio de Janeiro state) with 752
tweets. The most geotagged urban park was Américo Renné Giannetti Municipal Park
(Minas Gerais state) with 2,615 tweets, followed by Ponte dos Bilhares municipal Park
(Manaus state) with 1,101 tweets and Flamboyant Municipal Park (Goiás state) with 1,033
tweets.

Fig 6. Geographical distribution of the tweets posted about Brazilian protected areas.
Choropleth map representing the number of tweets posted about Brazilian protected areas in

51
Brazil and worldwide from 2011-2020. Locations are extracted from geotagged tweets and Twitter
user profiles. The map was cropped above Alaska to enhance the visibility of the other countries.

4.4.4 Which protected areas generate most public interest and engagement?

Based on the analysis of tweet content, national parks were the most posted
category of PAs on Twitter, followed by state and municipal parks (Fig 7A). The iconic
Iguaçu National Park received over 21,700 tweets, and Tijuca National Park received
9,507 tweets. Tourism emerged as the primary topic tweeted about among the top 30
most-tweeted parks; seven out of the ten most visited national parks in Brazil [52] were
among the top 12 most-tweeted about parks. We hypothesized that PAs that generated
high public interest, with numerous posts about them, also garnered high engagement
(likes + retweets). When evaluating the average engagement associated with PAs,
besides national parks, other PA categories surfaced, including environmental protection
areas, national forests, and biological reserves (Fig 7B).

Fig 7. Brazilian protected areas that have generated the most public interest and
engagement. (A) Lollipop plot illustrating the volume of tweets about Brazilian protected areas
during the period from 01 January 2011 to 31 December 2020. (B) Lollipop plot illustrating the
average of engagement in the tweets related to Brazilian protected areas. The top 30 tweets were
illustrated in ascending order and grouped by level of protection: federal, state, and local.

In addition to Iguaçu National Park, our results indicate that the most tweeted PAs
did not achieve the highest levels of engagement (see comparison in Figs 7A and 7B).
Instead, tweets about conflicts, rather than tourism activities, generated the highest
engagement, and these tweets were more frequently associated with other PA
designations. Upon analysing the five most engaged PAs on Twitter, the tweets with the
highest engagement included discussions about fires, jaguar deaths, and administrative
abuses, among the most frequent topics.

53
4.4.5 Content analysis of tweets

A broader look at the topics discussed in the full dataset revealed somewhat similar
patterns. Our cluster analysis of the most commonly featured keywords in all tweets
identified five thematic categories. Based on the most frequent words in each cluster
represented in the dendrogram (Fig 8), we characterised the main topics of discussion as
follows:
(i) Fires: this cluster included words associated with fires in PAs;
(ii) Management of protected areas: themes related to environmental crimes; PADDD
(downgrading, downsizing, and degazettement) events, mainly focused on the size of
PAs, and educational campaigns;
(iii) Nature protection: with regard to the creation of new PAs;
(iv) Nature appreciation: this cluster included words associated with sharing on social
media and natural monuments;
(v) Visitation: where the names of national parks, and urban parks were mentioned the
most.

Fig 8. Most published content related to Brazilian protected areas. The contents were published in the period 01 January 2011 to
31 December 2020. Each branch in the dendrogram represents a different cluster, grouped according to the similarity of words and
themes.

4.5 Discussion
We assessed Twitter users' interest in Brazilian PAs over a ten-year time scale
(2011-2020) using metrics of content and engagement, including likes and retweets in the
Portuguese language. Our results suggest that social media content related to PAs
remained broadly stable throughout the studied period, though engagement with this
content was low until 2016, at which time engagement began to increase. Engagement
grew particularly steeply after 2018, following a remarkable increase in likes and retweets
related to Brazilian PAs, even though the number of users posting on this theme remained
relatively stable over time. There are at least two factors that could be driving this pattern:
First, in 2016 Twitter introduced an algorithm to customise the content that users can
access on their profile’s timeline [53]. Based on accounts users have chosen to follow and
posts interacted with, the algorithm presents users with a series of recommendations. This
new personalization algorithm amplifies certain content while reducing the visibility of
other posts [54]. In this way, people sympathetic to environmental themes will be more
likely to be presented with content related to nature (an “echo chamber” effect), such as
PAs, and may more readily engage with posts (tweets and retweets) that in the past may
not have been easily encountered. In addition, users who carried the verification seal on
Twitter, identifying them as figures of public interest, may have experienced a significant
increase in visibility and influence within the platform after 2016 (Fig 5). This change can
be attributed to the way Twitter's algorithm has optimised the dissemination of tweets from
these verified users, especially by making their tweets become viral.
A second reason for the observed engagement trend may be related to the change
in Brazilian federal government that took place at the end of the same year, and the
increase in public discourse about the environment and environmental policy that started
during the political campaign that preceded the election and continued after it took place.
In 2019, the recently elected Bolsonaro government made a number of highly
controversial decisions to backtrack on environmental policies [55]. Specifically, the
incoming government made pledges to halt the expansion of the PA system and to make
environmental licensing more flexible [56] leading to a wide scale mobilisation of the
environmental movement in Brazil. The resulting conflicts of perspectives and attitudes

56
generated considerable discussion on Twitter and beyond. This, combined with Twitter's
new algorithm, may have prompted concerned citizens to express their opinions (agreeing
or disagreeing with government policy) by engaging with content related to what was
happening to Brazilian PAs.
Notwithstanding the highly polarised debate around Brazilian PAs and
improvements in personalization of Twitter feeds, our results indicate that the volume of
posts made by a single user was positively but weakly correlated with levels of
engagement. Also, posts by environmental NGOs, official agencies and celebrities had a
much greater level of engagement, as has been found in other studies [25]. As illustrated
in Figure 5, the growth trajectory of engagement regarding PAs after 2016 is similar
between the types of users (verified and unverified). Nevertheless, it is verified users who
play a key role in feeding interest and engagement in this subject, due to the substantial
number of followers, running into thousands or even millions, who often express
appreciation through likes, comments and retweets - a result which, perhaps
unsurprisingly, highlights the ability of celebrities to captivate the Twittersphere's attention.
Celebrities have the ability to attract people's attention [57] and have a long and complex
history of environmental engagement and activism (reviewed in [58]). There were several
artists and politicians who only contributed with one or two posts about Brazilian PAs in
our database, but due to their enormous number of followers, generated very high levels
of public engagement. This further confirms the importance and power of celebrities in
raising awareness of environmental issues, promoting science and environmental
conservation [57,58]. More generally, recent research has highlighted the existence of
different personas engaged in environmental discourses on social media [59], and our
results suggest a similar result for PA discussions. Exploring in greater detail which
personas drive and shape PA discussions on social media can help better understand
public discussions around this topic.
We found a relatively low number of georeferenced posts, possibly due to the
decision to collect only Portuguese language tweets. In Brazil, the most populated states
generated the most tweets. However, Iguaçu National Park, despite not being located in
a highly populated state, had the highest volume for an individual PA. This supports the
findings of a recent study of internet salience that found that Iguaçu National Park was the

57
most mentioned Brazilian PA on the national and global internet [26]. This may be due to
its beauty, iconic status and large annual number of visitors [51]. Iguaçu is the second
most visited PA in the country, with excellent visitation structure and is located on the
border between Brazil and Argentina - the second ranked country with the most georeferenced tweets about Brazilian PAs.
Another surprising finding was the high number of urban parks georeferenced in
the tweets. Urban parks are open green areas that can perform ecological, landscape and
recreational functions [60]. Significantly, they are not classified as PAs according to
Brazilian environmental policy. This observation highlights the importance of more
effective communication regarding PAs and the need for awareness about their functions
and values. Despite not being classified by the Brazilian environmental policy as PAs [3],
such parks can provide important physical, social and health benefits for urban residents
[61]. They are often the gateway for people's first contact with nature, and provide
opportunities to escape from the stressful pace of the city [62]. Being located in urban
areas, these parks are more easily accessed when compared to other PAs and tend to
have better mobile phone signal coverage. This potentially allows more real-time
interaction with social media, contributing to the high volume of georeferenced postings.
As an example, the Américo Renné Giannetti Municipal Park in the Minas Gerais state
had the second highest volume of georeferenced posts in the entire database.
The use of social media data has recognized potential for improving knowledge
and monitoring tourism interest in PAs [12,13,63]. Our study suggests that the most
officially visited parks were also the most popular subjects for Twitter posts. Such a result
matches with previous research that used different research platforms such as Google
Trends [26], OpenStreetMap (OSM), and Wikipedia [64,65] and social media platforms
such as Instagram, Flickr, and Twitter [66,67]. However, two highly-visited national parks
drew attention for not being among the 30 most posted-about on Twitter: Jericoacoara
National Park and the Fernando de Noronha Marine National Park. Similar discrepancies
between visitation and tweets were observed in research on Nepalese Parks, due to
greater local visitation and Twitter use restrictions [25]; factors that are less likely to
account for the discrepancies observed in the current study. These are more likely caused
by: (i) intrinsic biases in data collection. A limitation of nearly all textual content studies on

58
the internet is the problem of synonyms [68] - alternative names or spellings for the
represented entity. In the current study we used search strings that contained the full
names of PAs, and may therefore have missed many tweets that used colloquial names.
In the cases of Jericoacoara National Park, many Brazilians refer to the Park as "Jeri" and
Fernando de Noronha Marine National Park as "Noronha"; (ii) longer names tend to be
less popular on social media platforms [69] and are more likely to contain spelling errors.
Abbreviations and misspellings can restrict the search for PAs, however, currently, this is
the only viable and standardised way to attach a name to the PA designation at scale [26]
and; (iii) Both Jericoacoara and Fernando de Noronha National Parks may have
communication gaps regarding their identity as National Parks. The two parks are located
in places (municipality and island, respectively) that have the same names as the Park
(Jijoca de Jericoacoara municipality and Fernando de Noronha Island). This could
confuse visitors or simply lead them to communicate about the place more broadly rather
than the park specifically which, again, could lead to underrepresentation in our database
due to our chosen search method.
The Park category was the most tweeted category, especially National Parks. The
National Park has the primary objective of preserving natural ecosystems of great
ecological significance and scenic beauty, enabling scientific research and the
development of environmental education and interpretation activities, nature-based
recreation, and ecotourism [3]. It is almost certainly that this high public interest, reflected
in the substantial number of Twitter posts, is attributed to the fact that national parks
incorporate recreation as a significant component of their objectives [70]. Recreation
activities are strongly linked to public interest and support for PAs. Moreover, National
Parks are the oldest and most visited PA category [71], and many were protected to
preserve the scenic beauty of their unique landscapes along with other relevant
biophysical assets [72]. In addition, in Brazil, National Parks have larger use concessions
and receive more tourists, raising more resources to invest in the media. However, we
noticed that a large part of our dataset contained municipal parks that are not currently
present in the CNUC database. The most general explanation for this, according to
government environmental analysts, is that the registration of municipal units is at the
discretion of municipal environmental agencies and is not mandatory. However,

59
registration in the CNUC is used to verify the criteria for the allocation of funds from federal
public policies, such as those from federal environmental compensation. Such funds are
destined exclusively for PAs recognized by CNUC as belonging to the SNUC. Our findings
indicate that many parks currently unlisted by the CNUC are generating considerable
value for society [73] and for biodiversity [74] and should therefore be a valid target for
financial investment and improved environmental management [75].
A diversity of topics was found on tourist experiences and nature appreciation (Fig
8), corroborating the importance of PAs in bringing people closer to nature and the
physical and psychological benefits they provide [11]. Unlike other content-based
research on PAs, our findings did not identify topics related to iconic animals [12,25].
Although Brazil is renowned for its megadiverse ecosystems, offering the opportunity to
witness wildlife in PAs such as west indian manatee (Trichechus manatus), river dolphins
(Inia geoffrensis), jaguar (Panthera onca), and the largest terrestrial mammal in Brazil, the
tapir (Tapirus terrestris) [76–78], it's important to note that many Brazilian National Parks
were established with the primary goal of conserving scenic landscapes (as referenced in
Article 11. [3]), rather than focusing specifically on iconic species. Furthermore, the type
of social media and PA category may influence the results related to the topics that are
chosen as symbols and the experiences that are chosen to be shared. Besides the topics
related to tourism experiences, most of the topics that generate high levels of public
interest were related to management actions and conflicts, such as fires, reduction in the
size of PAs, and possible environmental crimes. This result corroborates our findings
related to engagement with posts (Fig 7B). These results highlight the potential use of
social media to monitor PAs in terms of cultural value generated (e.g., for tourists) or public
discontent related to management decision-making [14,79]. In the context of the spread
of misinformation on Twitter, it is important to recognise that false messages may have
influenced user engagement in relation to PAs, despite our observation that the presence
of bots did not have a significant impact in our study (only 2.61% of the tweets were
duplicates and could potentially be attributed to bots). Previous research [80] indicates
that bots play a role in the dissemination of true and false information, but it is human
action that amplifies false news more quickly and extensively, as they tend to arouse a
greater sense of novelty and excitement in people.

60
It is well known that the dissemination of incorrect information can result in the
delegitimisation of science, the minimisation of real threats to conservation and the
generation of polarisation about the importance of protected areas. In this sense, there is
a clear need for additional research that focuses on textual and behavioural content
analysis in order to understand whether the issues that lead people to engage with
environmental issues are false or true. These studies can shed light on the dynamics
behind the spread of misinformation, as well as on effective communication strategies to
combat its detrimental effects on discussions about protected areas and other
environmental issues. Better understanding of human-PA interactions can also be used
to inform strategies to attract societal support for conservation and PAs, to improve
conservation planning and management, and to tailor communication strategies
[13,15,37].
4.6 Conclusions
Overall, our study adds to the rapidly growing literature on the use of culturomic
metrics for monitoring human-nature interactions at large spatio-temporal scales [17].
Nevertheless, Twitter data has a few limitations that limit its suitability as a stand-alone
monitoring tool. First, Twitter corpora contain high levels of slang, colloquialisms,
acronyms and emoticons that can be challenging for data collection and analysis. Second,
social networks are not necessarily representative of PA users or of wider society [25].
Finally, not all users enable the geolocation on their devices, limiting this type of data [42].
Thus, all social media data requires extensive cleaning and critical analysis [15]. Despite
such limitations, the enormous volumes of data generated by social media platforms such
as Twitter mean that their analysis can certainly generate insights into the perceptions of
non-social media users [18] and, critically, allows the evaluation of macroscopic patterns
of human-nature interactions [81]. Acknowledging both the limitations and potentials
inherent in utilising social media as a means of investigation, we propose the integration
of various online data sources, such as Wikipedia, Instagram, Facebook, and offline
methodologies like questionnaires. This comprehensive approach aims to reinforce the
discourse surrounding the attained outcomes and enhance the depth of analysis.
Accepting the intrinsic limitations of our data, our analysis of 10 years of Twitter discourse

61
about Brazilian PAs indicates several areas of policy that could be improved. For example,
it is clear that official communication about PAs could be improved, from providing and
disseminating basic information about some parks (Fig 5), to the way government
agencies responsible for managing PAs communicate their decisions and campaigns (Fig
7). The lack of public engagement with conservation is often attributable to ineffective
communication from scientists and decision-makers [21]. In this sense, government
agencies and environmental NGOs could use data from social media to create better and
more powerful awareness campaigns, potentially in partnership with celebrities and other
social media ‘influencers’ (Fig 5). Furthermore, understanding online perceptions and
public interest can lead to identifying topics related to conservation actions (Fig 8), where
managers can pinpoint gaps and enhance strategies to maximise positive outcomes and
minimise negative impacts. In a broader context, our study confirms that the use of content
and engagement metrics on social media, combined with other monitoring tools such as
environmental data on water quality, air, biodiversity, along with information from field
monitoring and traditional media, has significant potential to enable early warnings to
identify emerging conflicts in PAs (Fig. 7B) and to identify public interest in these areas
(Fig. 7A).
4.7. Acknowledgments
We are grateful to all the LACOS21 colleagues, who contributed with insights to
this work. We also appreciate the comments made by the editor, Joseph Millard and one
anonymous reviewer that greatly contributed to improve the original manuscript.

REFERENCES
1. Watson JEM, Dudley N, Segan DB, Hockings M. The performance and potential of
protected areas. Nature. 2014;515: 67–73. doi:10.1038/nature13947
2. Maretti C, Catapan M, Abreu MJ, Oliveira J. Áreas protegidas: definições, tipos e
conjuntos – reflexões conceituais e diretrizes para a gestão. 2012.
3. BRASIL. Lei No 9.985, de 18 de julho de 2000. institui o Sistema Nacional de
Unidades de Conservação da Natureza e dá outras providências. Law. 2000.

62
4. BRASIL. Decreto n°. 5758, de 13 de abril de 2006. 2006.
5. Engen S, Hausner VH, Gurney GG, Broderstad EG, Keller R, Lundberg AK, et al.
Blue justice: A survey for eliciting perceptions of environmental justice among coastal
planners’ and small-scale fishers in Northern-Norway. PLoS One. 2021;16: 1–20.
doi:10.1371/journal.pone.0251467
6. Gerhardinger LC, Godoy EAS, Jones PJS, Sales G, Ferreira BP. Marine protected
dramas: The flaws of the Brazilian national system of marine protected areas. Environ
Manage. 2011;47: 630–643. doi:10.1007/s00267-010-9554-7
7. Bragagnolo C, Gamarra NC, Malhado ACM, Ladle RJ. Proposta Metodológica para
Padronização dos Estudos de Atitudes em Comunidades Adjacentes às Unidades de
Conservação de Proteção Integral no Brasil. Biodiversidade Brasileira. 2016;6(1): 190–
208.
8. Silva JMC da, Dias TCA de C, Cunha AC da, Cunha HFA. Funding deficits of
protected areas in Brazil. Land use policy. 2021;100.
doi:10.1016/j.landusepol.2020.104926
9. Bernard E, Penna LAO, Araújo E. Downgrading, downsizing, degazettement, and
reclassification of protected areas in Brazil. Conservation Biology. 2014;28: 939–950.
doi:10.1111/cobi.12298
10. Jepson PR, Caldecott B, Schmitt SF, Carvalho SHC, Correia RA, Gamarra N, et al.
Protected area asset stewardship. Biol Conserv. 2017;212: 183–190.
doi:10.1016/j.biocon.2017.03.032
11. De Haan FJ, Ferguson BC, Adamowicz RC, Johnstone P, Brown RR, Wong THF.
The needs of society: A new understanding of transitions, sustainability and liveability.
Technol Forecast Soc Change. 2014;85: 121–132. doi:10.1016/j.techfore.2013.09.005
12. Hausmann A, Toivonen T, Fink C, Heikinheimo V, Kulkarni R, Tenkanen H, et al.
Understanding sentiment of national park visitors from social media data. People and
Nature. 2020; pan3.10130. doi:10.1002/pan3.10130
13. Becken S, Stantic B, Chen J, Alaei AR, Connolly RM. Monitoring the environment
and human sentiment on the Great Barrier Reef: Assessing the potential of collective
sensing. J Environ Manage. 2017;203: 87–97. doi:10.1016/j.jenvman.2017.07.007
14. Ladle RJ, Correia RA, Do Y, Joo GJ, Malhado ACM, Proulx R, et al. Conservation
culturomics. Front Ecol Environ. 2016;14: 269–275. doi:10.1002/fee.1260
15. Toivonen T, Heikinheimo V, Fink C, Hausmann A, Hiippala T, Järv O, et al. Social
media data for conservation science: A methodological overview. Biol Conserv.
2019;233: 298–315. doi:10.1016/j.biocon.2019.01.023

63
16. Minin E Di, Tenkanen H, Toivonen T. Prospects and challenges for social media
data in conservation science. Front Environ Sci. 2015;3: 1–6.
doi:10.3389/fenvs.2015.00063
17. Correia RA, Ladle R, Jarić I, Malhado ACM, Mittermeier JC, Roll U, et al. Digital data
sources and methods for conservation culturomics. Conservation Biology. 2021;35:
398–411. doi:10.1111/cobi.13706
18. Ceron A, Curini L, Iacus SM, Porro G. Every tweet counts? How sentiment analysis
of social media can improve our knowledge of citizens’ political preferences with an
application to Italy and France. New Media Soc. 2014;16: 340–358.
doi:10.1177/1461444813480466
19. Howison, J., Wiggins, A., & Crowston, K. (2011). Validity issues in the use of social
network analysis with digital trace data. Journal of the Association for Information
Systems, 12(12), 2.
20. Lamb CT, Gilbert SL, Ford AT. Tweet success? Scientific communication correlates
with increased citations in Ecology and Conservation. PeerJ. 2018;2018.
doi:10.7717/peerj.4564
21. Papworth SK, Nghiem TPL, Chimalakonda D, Posa MRC, Wijedasa LS, Bickford D,
et al. Quantifying the role of online news in linking conservation research to Facebook
and Twitter. Conservation Biology. 2015;29: 825–833. doi:10.1111/cobi.12455
22. Otsuka R, Yamakoshi G, Id RO, Yamakoshi G. Analyzing the popularity of YouTube
videos that violate mountain gorilla tourism regulations. PLoS One. 2020;15: 1–20.
doi:10.1371/journal.pone.0232085
23. Fink C, Hausmann A, Di Minin E. Online sentiment towards iconic species. Biol
Conserv. 2020;241. doi:10.1016/j.biocon.2019.108289
24. Wu Y, Xie L, Huang SL, Li P, Yuan Z, Liu W. Using social media to strengthen public
awareness of wildlife conservation. Ocean Coast Manag. 2018;153: 76–83.
doi:10.1016/j.ocecoaman.2017.12.010
25. Bhatt P, Pickering CM. Public Perceptions about Nepalese National Parks: A Global
Twitter Discourse Analysis. Soc Nat Resour. 2021;34: 683–700.
doi:10.1080/08941920.2021.1876193
26. Correia RA, Jepson P, Malhado ACM, Ladle RJ. Culturomic assessment of Brazilian
protected areas: Exploring a novel index of protected area visibility. Ecol Indic. 2018;85:
165–171. doi:10.1016/j.ecolind.2017.10.033
27. Statista. Worldwide digital population as of January 2021. 2021 [cited 9 Mar 2022].
Available: https://www.statista.com/statistics/617136/digital-population-worldwide/.

64
28. Statista. Social media usage in Brazil – statistics & facts. In: 2022 [Internet]. [cited 19
Jan 2022]. Available: https://lb-aps-frontend.statista.com/topics/6949/social-mediausage-in-brazil/#topicHeader__wrapper
29. Collins K, Shiffman D, Rock J. How are scientists using social media in the
workplace? PLoS One. 2016;11. doi:10.1371/journal.pone.0162680
30. Mohammadi E, Thelwall M, Kwasny M, Holmes KL. Academic information on Twitter:
A user survey. PLoS ONE. Public Library of Science; 2018.
doi:10.1371/journal.pone.0197265
31. Kirilenko AP, Stepchenkova SO. Public microblogging on climate change: One year
of Twitter worldwide. Global Environmental Change. 2014;26: 171–182.
doi:10.1016/j.gloenvcha.2014.02.008
32. Mittermeier RA, Da Fonseca GAB, Rylands AB, And ‡, Brandon K. A Brief History of
Biodiversity Conservation in Brazil. Conservation Biology. 2005;19: 601–607.
33. Prates PL, Irving AMA. Conservação da biodiversidade e políticas públicas para as
áreas protegidas no Brasil: desafios e tendências da origem da CDB às metas de Aichi.
Revista Brasileira de Políticas Públicas. 2015. 10.5102/rbpp.v5i1.3014
34. Medeiros R, Araújo FFDS. Dez anos do sistema nacional de unidades de
conservação da natureza: lições do passado, realizações presentes e perspectivas para
o futuro. 2011.
35. MMA. Cadastro das Unidades de Conservação – CNUC. 2021 [cited 19 Dec 2021].
Available: https://dados.gov.br/dataset/unidadesdeconservacao/resource/baf254485064-4ece-9a0e-d778b0eca542?inner_span=True
36. Jenkins CN, Joppa L. Expansion of the global terrestrial protected area system. Biol
Conserv. 2009;142: 2166–2174. doi:10.1016/j.biocon.2009.04.016
37. Mascia MB, Pailler S. Protected area downgrading, downsizing, and degazettement
(PADDD) and its conservation implications. Conserv Lett. 2011;4: 9–20.
doi:10.1111/j.1755-263X.2010.00147.x
38. Ladle RJ, Jepson P, Correia RA, Malhado ACM. The power and the promise of
culturomics. Front Ecol Environ. 2017;15: 290–291. doi:10.1002/fee.1506
39. Barrie C, Ho J. academictwitteR: an R package to access the Twitter Academic
Research Product Track v2 API endpoint. J Open Source Softw. 2021;6: 3272.
doi:10.21105/joss.03272
40. Feinerer I; Hornik K, Meyer D. Text Mining Infrastructure in R. J Stat Softw. 2008;25.
41. Harrell Jr F. Hmisc: Harrell Miscellaneous. 2023.

65
42. Mangiafico S. rcompanion: Functions to Support Extension Education Program
Evaluation. New Jersey; 2023.
43. Otero P, Gago J, Quintas P. Twitter data analysis to assess the interest of citizens
on the impact of marine plastic pollution. Mar Pollut Bull. 2021;170: 112620.
doi:10.1016/j.marpolbul.2021.112620
44. python-visualization. Folium. 2020. Available: https://pythonvisualization.github.io/folium/
45. Kassambara A. Multivariate Analysis I Practical Guide To Cluster Analysis in R
Unsupervised Machine Learning. 2017. Available: http://www.sthda.com
46. Maechler, M; Rousseeuw, P; Struyf, A; Hubert, M; Hornik K. cluster: Cluster Analysis
Basics and Extensions. 2022. Available: https://cran.r-project.org/package=cluster.
47. Vieira, F.A.S., Santos, D.T.V., Bragagnolo, C., Campos-Silva, J.V., Correia, R.A.H.,
Jepson, P., Malhado, C.M., Ladle, R.J., 2021. Social media data reveals multiple cultural
services along the 8.500 kilometers of Brazilian coastline. Ocean & Coastal
Management, 214, 105918
48. Wickham, H.; François, R.; Henry, L.; Müller K. dplyr: A Grammar of Data
Manipulation. 2021. Available: https://cran.r-project.org/package=dplyr
49. Wickham H. Elegant Graphics for Data Analysis: ggplot2. Applied Spatial Data
Analysis with R. 2008.
50. Sievert C. Interactive Web-Based Data Visualization with R, plotly, and shiny. 2020.
doi:https://doi.org/10.1201/9780429447273
51. Kassambara, A.; Mundt F. factoextra: Extract and Visualize the Results of
Multivariate Data AnalysesNo Title. 2020. Available: https://cran.rproject.org/package=factoextra
52. Instituto Chico Mendes de Conservacao da Biodiversidade. Painel Dinâmico de
Informações. Uso Público e Turismo. 2021 [cited 9 Mar 2022]. Available:
http://qv.icmbio.gov.br
53. @mjahr. Never miss important tweets from people you follow. 2016 [cited 3 Mar
2022]. Available: https://blog.twitter.com/en_us/a/2016/never-miss-important-tweetsfrompeople-you-follow
54. Huszár F, Ktena SI, O’brien C, Belli L, Schlaikjer A, Hardt M. Algorithmic
amplification of politics on Twitter. Proceedings of the National Academy of Sciences.
2022;119. doi:10.1073/pnas.2025334119/-/DCSupplemental
55. Fearnside PM. Setbacks under President Bolsonaro: A Challenge to Sustainability in
the Amazon. Sustentabilidade International Science Journal. 2019;1: 38–52.

66
56. Ferrante L, Fearnside PM. Brazil’s new president and “ruralists” threaten Amazonia’s
environment, traditional peoples and the global climate. Environmental Conservation.
Cambridge University Press; 2019. doi:10.1017/S0376892919000213
57. Craig G. Celebrities and Environmental Activism. Media, Sustainability and Everyday
Life. Palgrave Macmillan UK; 2019. pp. 135–163. doi:10.1057/978-1-137-53469-9_6
58. Brockington D. Celebrity and the Environment: Fame, Wealth and Power in
Conservation. London: Zed Books; 2009.
59. Chang CH, Armsworth PR, Masuda YJ. Twitter data reveal six distinct environmental
personas. Front Ecol Environ. 2022;20: 481–487. doi:10.1002/fee.2510
60. Nielsen AB, van den Bosch M, Maruthaveeran S, van den Bosch CK. Species
richness in urban parks and its drivers: A review of empirical evidence. Urban Ecosyst.
2014;17: 305–327. doi:10.1007/s11252-013-0316-1
61. Özgüner H. Cultural differences in attitudes towards urban parks and green spaces.
Landsc Res. 2011;36: 599–620. doi:10.1080/01426397.2011.560474
62. Chiesura A. The role of urban parks for the sustainable city. Landsc Urban Plan.
2004;68: 129–138. doi:10.1016/j.landurbplan.2003.08.003
63. Hausmann A, Toivonen T, Slotow R, Tenkanen H, Moilanen A, Heikinheimo V, et al.
Social Media Data Can Be Used to Understand Tourists’ Preferences for Nature-Based
Experiences in Protected Areas. Conservation Letters. Wiley-Blackwell; 2018.
doi:10.1111/conl.12343
64. Guedes-Santos J, Correia RA, Jepson P, Ladle RJ. Evaluating public interest in
protected areas using Wikipedia page views. J Nat Conserv. 2021;63.
doi:10.1016/j.jnc.2021.126040
65. Levin N, Lechner AM, Brown G. An evaluation of crowdsourced information for
assessing the visitation and perceived importance of protected areas. Applied
Geography. 2017;79: 115–126. doi:10.1016/j.apgeog.2016.12.009
66. Fisher DM, Wood SA, Roh YH, Kim CK. The geographic spread and preferences of
tourists revealed by user-generated information on jeju island, south korea. Land
(Basel). 2019;8. doi:10.3390/LAND8050073
67. Tenkanen H, Di Minin E, Heikinheimo V, Hausmann A, Herbst M, Kajala L, et al.
Instagram, Flickr, or Twitter: Assessing the usability of social media data for visitor
monitoring in protected areas. Sci Rep. 2017;7. doi:10.1038/s41598-017-18007-4
68. Correia RA, Jepson P, Malhado ACM, Ladle RJ. Internet scientific name frequency
as an indicator of cultural salience of biodiversity. Ecol Indic. 2017;78: 549–555.
doi:10.1016/j.ecolind.2017.03.052

67
69. Zmihorski M, Dziarska-Palac J, Sparks TH, Tryjanowski P. Ecological correlates of
the popularity of birds and butterflies in Internet information resources. Oikos. 2013;122:
183–190. doi:10.1111/j.1600-0706.2012.20486.x
70. Dudley N. Guidelines for applying Protected Area Management Categories. IUCN,
Gland, Switzerland. IUCN; 2008. doi:10.1103/PhysRevB.38.10724
71. Figueiredo CCM. From paper parks to real conservation: case studies of national
park management effectiveness in Brazil. Ohio State University. 2007. Available:
https://etd.ohiolink.edu/apexprod/rws_etd/send_file/send?accession=osu1167587930&di
sposition=inline
72. Gamarra NC, Correia RA, Bragagnolo C, Campos-Silva JV, Jepson PR, Ladle RJ, et
al. Are Protected Areas undervalued? An asset-based analysis of Brazilian Protected
Area Management Plans. J Environ Manage. 2019;249: 109347.
doi:10.1016/j.jenvman.2019.109347
73. Prado DAR. Parque Municipal Flamboyant: apropriação e usos para lazer.
Universidade Federal de Goiás. 2012. Available:
https://repositorio.bc.ufg.br/tede/handle/tde/1880
74. Calheiros AR, Souza MA, Costa JG da, Araújo KD. Espécie invasora de bambu e
seus impactos sobre a qualidade do solo. Revista Ibero-Americana de Ciências
Ambientais. 2023;13: 63–73. doi:10.6008/cbpc2179-6858.2022.006.0006
75. Bragagnolo C, Correia RA, Gamarra NC, Lessa T, Jepson P, Malhado ACM, et al.
Uncovering assets in Brazilian national parks. J Environ Manage. 2021;287.
doi:10.1016/j.jenvman.2021.112289
76. Izidoro FB, Schiavetti A. Associated benefits of manatee watching in the Costa dos
Corais Environmental Protection Area. Front Mar Sci. 2022;9.
doi:10.3389/fmars.2022.1002855
77. Vidal MD, Santos MC, Jesus JS, C Santos PM, Jesus JS, Alves LCPS, Chaves,
MPSR. Ordenamento participativo do turismo com botos no Parque Nacional de
Anavilhanas, Amazonas, Brasil. Bol Mus Para Emílio Goeldi Cienc Nat. 2017.
78. Tortato FR, Ribas C, Concone HVB, Hoogesteijn R. Turismo de observação de
mamíferos no Pantanal. Boletim do Museu Paraense Emílio Goeldi - Ciências Naturais.
2022;16: 351–370. doi:10.46357/bcnaturais.v16i3.814
79. Almeida JAGR, Guedes-Santos J, Vieira FAS, Azevedo AK, Souza CN, Pinheiro BR,
et al. Public awareness and engagement in relation to the coastal oil spill in northeast
Brazil. An Acad Bras Cienc. 2022;94: 1–10. doi:10.37002/biobrasil.v12i2.2177
80. Vosoughi S, Roy D, Aral S. The spread of true and false news online. Science. 2018
Mar 9;359(6380):1146-1151. doi: 10.1126/science.aap9559. PMID: 29590045.

68
81. Ladle, RJ, Souza, CN, Correia R. Culturomics for (not against!) protected areas In.
Biol Conserv. 2021;256: 109197. doi:10.1016/j.biocon.2021.109015

69
5 USING SOCIAL MEDIA AND MACHINE LEARNING TO UNDERSTAND
SENTIMENTS TOWARDS BRAZILIAN NATIONAL PARKS4

Carolina Neves Souzaa*,+; Javier Martínez-Arribasb,+; Ricardo A. Correiac,d,e; João A. G.
R. Almeidaf; Richard Ladleb; Ana Sofia Vazb; Ana Cláudia M. Malhadoa

Affiliations:
a Institute of Biological and Health Sciences, Federal University of Alagoas, Av. Lourival

Melo Mota, s/n, Tabuleiro do Martins, 57072-90, Maceió, AL, Brazil
b CIBIO-InBIO, Research Centre in Biodiversity and Genetic Resources, University of

Porto, Campus de Vairão, 4485-661 Vairão, Portugal
c Biodiversity Unit, University of Turku, 20014 Turku, Finland
d

Helsinki Lab of interdisciplinary Conservation Science (HELICS), Department of

Geosciences and Geography, University of Helsinki, 00014 Helsinki, Finland
e

Helsinki Institute of Sustainability Science (HELSUS), University of Helsinki, 00014

Helsinki, Finland
f

Institute of Computing, Federal University of Alagoas, Av. Lourival Melo Mota, s/n,

Tabuleiro do Martins, 57072-90, Maceió, AL, Brazil

*Corresponding author:
Carolina Neves Souza.
Address: Institute of Biological and Health Sciences, Federal University of Alagoas, Av.
Lourival Melo Mota, s/n, Tabuleiro do Martins, 57072-90, Maceió, AL, Brazil.

4 Artigo aceito para publicação na revista Biological Conservation.

https://doi.org/10.1016/j.biocon.2024.110557

+ Shared first authors.

5.1 Abstract
Protected areas (PAs) play a vital role in the conservation of natural and cultural heritage
while supporting local livelihoods. However, in Brazil, where limited resources and poor
effectiveness lead to negative sentiments and are leveraged as criticism towards PAs, it
is necessary to better comprehend public perceptions of Brazilian PAs and identify the
key factors contributing to negative sentiments. Here, we use data from online discussions
about Brazilian national parks (NPs) on Twitter and sentiment analysis to explore this
question. We classified the sentiment of ~100,000 tweets collected over a twelve-year
period (2011 2022) using the BERTimbau Base model. We also performed a topic
modelling with the BERTopic model to identify prevalent subjects concerning Brazilian
NPs. We identified 18,388 (17.30 %) posts expressing negative sentiment towards NPs,
mostly associated with wildfires occurring between 2011 and 2017 and concerning
government decisions impacting conservation efforts after 2019. The results revealed six
prominent topics: (1) Wildfires; (2) Security; (3) Regulations; (4) Wildlife roadkill; (5)
Privatization; (6) Lack of financial resources, reflecting a diverse range of negative sentiments regarding the parks, surpassing isolated events. Furthermore, examining specific
topics on a per-park basis proved beneficial in identifying distinct issues and conflicts in
the five most tweeted NPs, facilitating targeted conservation actions. Using social media
data to better understand public perceptions of NPs can strengthen their management
and governance by reinforcing their conservation initiatives and enhancing visitor
experiences. Our findings underscore the value of sentiment analysis in identifying gaps
and driving improvements in the management of protected areas.
5.2 Introduction
Protected areas (PAs) are a key strategy for promoting the preservation of
biological resources and the sustainable use of natural benefits, including ecosystem
services and cultural practices (Maretti, et al., 2012; Watson et al., 2014). As a

71
megadiverse country, Brazil has a significant responsibility to protect its biodiversity
(Rylands and Brandon, 2005). According to the National Register of Conservation Units
(CNUC) of 2023, Brazil is home to a total of 2,859 protected areas (PAs), which
collectively cover an area of approximately 2,583,237 km2. Considering the Brazilian
continental area, 19.01% is protected by PAs; while considering the Brazilian exclusive
economic zone, 26.49% is being protected by PAs (CNUC, 2023). Brazilian PAs face
numerous biophysical and political challenges, from climate change (Soares-Filho et al.,
2010) to limited resources for management and monitoring (Silva et al., 2021).
Furthermore, they are often viewed as opportunity costs (obstacles to economic
development) by politicians and decision-makers (Ferreira et al, 2014). Indeed, the
downsizing, downgrading, and degazettement of PA’s (PADDD) have affected 72,892km2
of Brazilian PAs between 1981 and 2012 (Bernard et al., 2014). Protecting Brazilian PAs
from PADDD requires that negative attitudes are countered by fully demonstrating their
value to society (e.g., Jepson et al., 2017) and by showing politicians that they have broad
public support (Guedes-Santos et al. 2021).
Of all categories, national parks perhaps have the most potential to draw society's
attention to the importance of PAs because they are often popular (Correia et al., 2018)
and hold an iconic status by reconciling conservation goals with opportunities to engage
with natural outdoor scenic attractions that can inspire and captivate visitors (Dudley,
2008). Moreover, national parks (category II by the IUCN), being the oldest and most
visited protected area category, hold a strong connection to recreational activities and a
broad range of assets that are important to visitors and generate broad societal value
(Bragagnolo et al., 2021), which are closely tied to public interest and support for PAs.
However, national parks can also elicit opposing views among visitors, representing either
conservation spaces with management restrictions (Hausmann et al., 2020), or positive
places for nature-human interactions with psychological and physical benefits for wellbeing (De Haan et al., 2014). It is important to recognize that the importance of individual
sentiments is not only limited to transitory expressions of emotions such as joy, anger,
interest, sadness, and gratitude. Rather, people's feelings towards national parks may
lead to long-term impact on people’s behaviour over the environment (Fredrickson, 2001).

72
According to Lemberg (2010), perception encompasses how people sense,
mentally process, and respond to information derived from their surroundings. It is shaped
by sociodemographic characteristics, attitudes, and values, which have the potential to
directly impact the experience, satisfaction, and behaviours associated with protected
areas (Hoeffel et al., 2008; Rossi et al., 2015). Within this context, it is important to
acknowledge that negative sentiments towards protected areas can unfavourably
influence perceptions, thereby leading to directed behaviours and dissatisfaction. The
connection between how people feel (sentiments) and their perceptions is very significant
for conservation, as adverse sentiments can fundamentally shape how people interpret
and interact with the natural environment. A concrete example of this is the concession of
national parks, which can have negative consequences on local communities, such as
impacting service provision as observed in Tijuca national Park (Maciel, 2015), and the
distancing of neighbouring communities and low-income visitors from national parks
described in Queensland, Australia (Rossi et al., 2016).
Developing a better understanding of how people perceive national parks, evaluate
their experiences, and identifying what motivates or hinders visitation is therefore crucial
for more informed decision-making (Agyeman et al., 2019; Bragagnolo et al., 2016; Griggs
and Lacey, 2022; Rossi et al., 2016). For example, understanding the mobility patterns of
visitors to protected areas is essential for formulating conservation strategies (Kim et al.
2023). This proves crucial not only for local management, but also for global marketing
initiatives, given that visitors' activities can result in direct or indirect impacts on the
environment (Toivonen et al. 2019). This knowledge can help guide the efforts of
managers and policymakers towards a more effective management of PAs identifying
which aspects and events lead to a negative public attitude towards PAs and addressing
more effective strategies to manage tensions and promote changes in favour of
conservation (Hausmann et al., 2020; Hockings et al., 2006; Instituto Semeia, 2022).
One of the major challenges in building this knowledge is accessing comprehensive
sources of information that can indicate people's feelings/perceptions towards the national
parks. In this context, social media has emerged as a significant source of information.
Social media has become an integral part of modern daily lives, offering unprecedented

73
opportunities for communication, debate, and information sharing, and thus presents a
powerful platform to understand public perceptions and feelings (Sudhir and Suresh,
2021). A key advantage of analysing digital data lies in the vast volume of information
generated by people on a wide range of topics, including political preferences (Ceron et
al., 2014), customer satisfaction (Ahani et al., 2019), and nature conservation (Souza et
al., 2023; Ladle et al. 2021, Di Minin et al., 2015). Compared to other methods such as
survey-based questionnaires, social media data can be collected quickly, at a lower
financial cost, and on a larger geographical scale (Becken et al., 2017), and can therefore
complement other more targeted approaches. The increasing availability of big online
social media data, including text, images, and videos, represents valuable and
complementary information for researchers, conservation practitioners, and policymakers
to explore citizens' opinions about PAs and biodiversity conservation (Correia et al., 2021).
Indeed, the use of online data from social media platforms such as Twitter, Facebook or
Instagram has already been applied to measure public interest and to understand users’
perceptions about a broad range of conservation topics (Almeida et al., 2022; Fink et al.,
2020; Papworth et al., 2015; Tenkanen et al., 2017).
One exciting tool for analysing digital data about human-nature interactions is
sentiment analysis, which leverages the use of social media data to comprehend people's
values and emotions towards the natural environment (Drijfhout et al., 2016). Sentiment
analysis employs computational techniques to extract and evaluate opinions and
emotions related to a specific entity or topic (Serrano-Guerrero et al., 2015). It has been
extensively applied in marketing and other domains, and more recently it has been used
by conservation scientists to assess people's sentiments on a range of topics including
environmental management (Bhatt and Pickering, 2021), the impacts of tourism on wildlife
(Otsuka et al., 2020), and tourists’ preferences, experiences, and opinions (Hausmann et
al., 2020). Sentiment analysis employs natural language processing (NLP) through
machine learning algorithms that classify the content of textual data based on positive,
negative, or neutral perceptions (Liu, 2012). It can provide a mechanistic understanding
of how people perceive and feel about nature and conservation (Drijfhout et al., 2016),
including PAs. For instance, it can detect negative emotions, such as frustration or
disappointment, and can indicate the need for improvements in the visitation areas of

74
protected areas (Agyeman et al., 2019) and in management practices or communication
strategies. Controversial management practices, such as restrictions on the use of space
and action against illegal hunting (Lubbe et. al. 2019) have the potential to generate
dissatisfaction, lack of support and generate conflicts. Thus, sentiment analysis applied to
social media data related to PAs can help identify problems such as those mentioned
above, and can support decision-making and improve management strategies.
Despite the great potential of sentiment analysis to generate insights for
conservation, one current barrier to its wider application is the limited availability of tools
and methods for languages other than English (Kaity and Balakrisnhan, 2020). To our
knowledge, no previous studies have leveraged social media data to analyse sentiments
about Brazilian PAs and this is at least partly due to methodological challenges of working
in Portuguese. However, such information is crucial to assist in the management of the
vast and diverse network of Brazilian PAs, whose vast biodiversity and ecosystems have
increasingly come under attack due to an attempt to dismantle environmental and
conservation policies over the past decade (Bernard et al., 2014, Fearnside, 2019; Vale
et al 2021). Furthermore, improving the application of sentiment analysis to other
languages, such as Portuguese, can ensure the broader application of these methods.
Here, our main objective is to understand the people's perception about the Brazilian
national parks through social media and sentiment analysis in the Portuguese language.
Specifically, we aim to identify the key topics that contribute to the public perceptions
(positive, neutral or negative) of Brazilian national parks. In this context, negative
sentiments are especially important as the factors that promote them are likely to
represent the main challenges for PA management. To achieve this goal, we used natural
language processing approaches to classify the sentiments related to Brazilian national
parks. In doing so, we hope to demonstrate how sentiment analysis can assist in
identifying opportunities to improve the management of Brazilian PAs.

5.3 Material and Methods
The methodological development of this study is divided into four distinct parts. The
first section contextualises the research area, focusing on Brazilian national parks.

75
Subsequently, we outline the data collection process carried out on the social media
platform Twitter, as well as the procedures adopted for data cleaning and filtering. In the
third part, we elucidate the sentiment analysis method. This section addresses the
underlying concepts of sentiment analysis, the model employed for sentiment
classification, and the datasets used for model training. Finally, we describe the analyses
conducted based on the results obtained from the sentiment classification of tweets
related to Brazilian national parks.

5.3.1 Study area
Our study focussed on the category of Brazilian protected area with the highest
number of tweets in our dataset, the national parks. Brazil has a vast and diversified
network of protected areas that safeguard its natural heritage, biodiversity, and cultural
resources (BRASIL, 2000). Brazilian national parks fall within the group of strictly
protected areas and are designed to protect natural areas of exceptional beauty, diversity,
and ecological significance (BRASIL, 2000). National parks are managed by the Chico
Mendes Institute for Biodiversity Conservation (ICMBio), with the aim of preserving natural
ecosystems, protecting threatened species, and promoting scientific research and
environmental education. National parks also provide different cultural and social services
to human populations (Nabout et. al 2022), through contemplation, religious rituals and
recreation, for example, which gives visitors the opportunity to get to know Brazil's
stunning landscapes and wildlife, promoting the protection and conservation of these
areas.
In 2022, when this study was carried out, there were 74 officially recognized
national parks in Brazil, covering an area of over 268.037 km2. National parks represent
2.59% of all conservation areas in Brazil, covering 3.11% of the country's land area and
contributing only 0.09% of the protected areas in the marine environment (CNUC, 2023).
In addition to marine areas, these parks are distributed across all of Brazil's different
biomes, including the Caatinga, Cerrado, Atlantic Forest, Pampa and Pantanal (Fig. 1).

Fig. 1. Map of distribution of 74 Brazilian national parks. Spatial distribution of the 74 Brazilian
national parks located in their respective biomes.

5.3.2 Data Collection
Digital data for conservation culturomics analysis can be applied to different
sources of digital data, such as texts, videos and images (Correia et al., 2021). The data
mining techniques involved data collection, cleaning, processing and analysis (see Fig.
2). In this study, we collected textual content from Twitter posts, using the Twitter v2 API
associated with a Twitter's Academic access which allowed it to collect up to 10 million
tweets monthly, access older conversation histories, and apply more search filters than
the basic API. It's worth noting that the restrictions and availability of APIs for free usage
have been volatile recently, as the platform has undergone a policy change in data access
and is now referred to as “X”. The data mining code was developed using the Python
language program v.3.9 (http://www.python.org) and was based on the Full-ArchiveSearch API node. A query composed of 18 keywords in Portuguese language (See

77
supplementary material) was selected to extract tweets related to 11 categories of
protected areas in Brazil, including national parks (Souza et al., 2023). We collected
tweets posted between 01 January 2011 to 31 December 2022, as well as information on:
(i) users (author_id, name, username); (ii) date of publication; (iii) geographic data of
tweets; (iv) publication data (number of likes, retweets, replies and whether it is a reply to
another user); and (v) text, of each tweet whenever available.

Fig. 2. Methodological flowchart. Methodological flowchart from data collection to results.
The flowchart showed all the steps used during the research: data collection, data analysis, and
results.

The tweets were downloaded by year and compiled into a single CSV file for data
cleaning and filtering. First, a filter was applied to the geographical metadata provided by
Twitter, specifically in the 'country' and 'country_code' columns, to identify tweets
originating from countries other than Brazil. Next, a manual review of each originally
foreign tweet was carried out to exclude those whose textual content was not related to
Brazilian protected areas (Souza et al., 2023). Finally, we selected from the dataset the
tweets that fell into the category of national parks - the core of our study. The final validated
list contained a total of 106,240 tweets about Brazilian national parks. (See
https://github.com/CIBIO-TropiBIO/Sentiment-Analysis-Brazilian-National-Parks).

78
5.3.3 Sentiment analysis
Sentiment analysis, also known as opinion mining, is a field of study that analyses
people's expressed opinions, sentiments, appraisals, attitudes, and emotions towards
entities and their attributes in written text (Birjali et al., 2021). The entities can take the
form of products, services, organisations, individuals, events, issues, or topics (Liu, 2012).
In this study, we focus on Brazilian national parks as our entity. Sentiment is typically
categorised as either "positive", "neutral", or "negative" and is assessed based on the
literal meaning of the text. For instance, the sentence "Itatiaia national park: the centre of
the problems of the universe is there!" expresses a negative sentiment, while "I liked it
and I recommend Iguaçu national park to all my friends! =)" expresses a positive
sentiment, and "I'm at Tijuca national park." expresses a neutral sentiment.
Different models, such as Naive Bayes, Maximum Entropy, Support Vector
Machine (SVM), and Bidirectional Encoder Representations for Transformers (BERT),
have been used to perform sentiment classification on large Portuguese-language
datasets (Pereira, 2021). However, the use of BERT is one of the most widely used
models in research (Souza et al. 2020) and although it already presents a multilingual
natural language processing model, the development of a monolingual approach such as
BERTimbau can be effective in training pre-trained language models, especially in
languages with few annotated resources, as in the case of Brazilian portuguese. These
models replicate the perception of human operators who manually classify the data, based
on statistical and structural patterns in the texts. In the case of BERT, this classification
can be better developed, since the model was created to have a greater sense of context
and language flow than one-way language models because it is trained bidirectionally.
Bidirectional training refers to an approach in which a language model is trained to
understand the context of a word or token not only by considering previous words, but
also subsequent words in a text. This approach favours the understanding of perceptions
written in text because they tend to better understand the context of texts and their
possible ambiguities (Souza et. al 2020). In this sense, we analysed the polarity of
people's expressions of opinions, sentiments, appraisals, attitudes, and emotions towards
the Brazilian national parks, based on a sentiment analysis of "positive", "neutral", or

79
"negative" meanings of the collected tweets. To do so, we adapted a pre-trained
BERTimbau Base model, due to its performance compared to other natural language
processing models and its focus on Brazilian portuguese language (Souza et al., 2020;
see Supplementary Material for details).
Our study mined over 100,000 tweets about Brazilian national parks. Although
sentiment analysis models perform better with data that is similar in terms of size, style,
and text type (Mozetič et al., 2016), good results can be achieved by using data of different
nature through appropriate preprocessing tailored to the form of the target data. In a first
attempt to classify the sentiments of the texts in our dataset, we used a corpus of tweets
with all three sentiments (Portuguese tweets for sentiment analysis, 2018). However, due
to limited information, the training of this first model led to most predictions being classified
as "neutral", resulting in errors when classifying "positive" and "negative" tweets about the
national parks. Therefore, in our second attempt, we decided to train a model using
another corpus that had also already been categorised into different types of sentiment
(positive, negative, or neutral) - the dataset containing over 200,000 user opinions in
Brazilian portuguese about products sold online to accomplish this task (B2W-Reviews01,
2018; Corpus Buscapé, 2013).
We used data similar to another study by Avanço and Nunes (2014), properly
preprocessed to better adapt these data to the target tweets, as it produced the most
accurate predictions for negative tweets which are the main focus of this work. This
opinion corpus has a "classification" field with scores ranging from 1 to 5, with 1 being the
most negative and 5 being the most positive. For our classification task, we transformed
scores 1 and 2 into "negative", 3 into "neutral", and 4 and 5 into "positive". To avoid
possible biases arising from different frequencies of occurrence between the categories,
we configured the algorithm to take this divergence into account. During the preprocess
of training, we assigned a lower weight to errors occurring in the more predominant
categories and a higher weight to errors in the rarer categories - what we can call
weighting between the classes.
To assess the accuracy of our automated classification, we manually and
independently annotated a random sample of 2,000 tweets from our dataset into three

80
classes of sentiment polarity. We then compared the manual annotation with the predicted
classes, which confirmed the second model's classification as the most reliable for
sentiment classification in our Brazilian national park database. The performance
evaluations that we employed to assess our model's performance encompassed accuracy
and F1-score. Accuracy means the proportion of true positives and true negatives divided
by the total number of predictions, or in other words, how correctly it predicts the
categories. This metric is applicable when all the classes are of equal importance, but if
there are a different number of observations for each category, it is possible to achieve
great accuracy by making all our predictions from the majority class, which is somewhat
illusory. For this reason, we also use the F1 score, which is a harmonic mean between
precision and recovery, offering a robust assessment for instances of misclassification,
i.e., false positives and false negatives (Capellaro, 2021), or a better metric of incorrectly
classified cases.
5.3.4 Time series analysis
We carried out a time series analysis of the number of daily negative and nonnegative tweets in order to identify the main events that generated the peaks in tweets
about Brazilian national parks. To do this, we summarised the number of tweets by
grouping them by date and year (2011-2022). We then generated a line graph describing
the changes in the number of tweets over the years. Using the plotly package (Sievert,
2020) in the R programming language (Team R, 2017), we interactively identified with the
graph which days had the highest number of tweets published, and the corresponding
event that potentially triggered these publication peaks was determined based on a search
of the textual content in our dataset. All analyses were carried out in the R programming
language (Team R, 2017), using the dplyr package (Wickham et al., 2021) for data
processing and the ggplot2 package (Wickham, 2008) for visual representations of the
data.

81
5.3.5 Topic modelling analysis
BERTopic is a Python library for natural language processing topic modelling that
combines transformer embeddings with clustering algorithms to identify topics in a corpus
of texts (Grootendorst, 2022). The BERTopic model supports over 50 languages and has
been compared to other models, such as LDA, for performing topic modelling on short
texts from social media platforms and has shown exceptional performance in extracting
topic representations (Egger and Yu, 2022).
Following the identification of the values associated with each term within the
topics, a comprehensive evaluation and inspection of the topics was conducted to detect
any potential content that might be misconstrued as a singular topic (See Table 1 in
results). This consideration, as noted by Egger and Yu (2022), highlights a potential
limitation of the model, particularly when dealing with extensive amounts of data for
analysis. Although BERTopic offers the advantage of leveraging domain-specific
knowledge to search for specific topics, as done in this study, this process can still be
considered exhaustive.
For the purpose of our study, two main steps were undertaken: (i) identification of
potential negative topics within our corpus, encompassing all Brazilian national parks, and
(ii) segregation of tweets specifically related to the five most tweeted parks in our dataset,
followed by clustering to discern the prominent negative topics associated with each
individual park. The reason for adopting this filter was to understand whether sentiment
analysis has the potential to identify specific topics that are particular to each park. To
achieve this, we performed the BERTopic model with the following hyperparameters:
•

For the Uniform Manifold Approximation and Projection (UMAP) algorithm we set
n_neighbors or the number of samples used during the manifold approximation to
15, n_components or the dimensionality that holds the most information possible to
5, min_dist to 0, in order to get more clustered embeddings and selected the cosine
metric to compute distances in high dimensional space.

•

For Hierarchical Density-Based Spatial Clustering of Applications with Noise
(HDBSCAN) we set the metric to Euclidean in order to compute distances in an array

82
and prediction_data to True to be able to apply to our dataset later, not just to fit the
model, for all the datasets, no matter what park the tweets are from. And we set the
min_cluster_size parameter or minimum size of the clusters depending on the
number of observations we have. The purpose is to reach a reasonable number of
topics and also that they contain coherent information to know what they are talking
about.
•

We set the parameter nr_topics to auto in order to focus on the interpretation of the
topics. Besides we use the function CountVectorizer with a list of portuguese stop
words and ngram_range between 1 and 2 n-gram words to be extracted, and the
function ClassTfidfTransformer in order to reduce the impact of the most frequent
words, also the MaximalMarginalRelevance function in order to limit the number of
duplicate words that we can find in each topic, and finally, the function
SentenceTransformer with the BERT-base-portuguese-cased model in order to use
the same embedding model for the negative tweets selected as in the previous
prediction step.
Due to the randomness of some parts of the BERTopic model, it is possible to

obtain a slight variation in the topics obtained in each execution of it. After several tests,
we identified, based on our knowledge, the topics that remained consistent across all
generated models.
5.4 Results

5.4.1 Public perceptions about Brazilian National Parks
We collected and analysed a total of 106,240 valid tweets about 74 Brazilian
national parks generated by 38,432 Twitter users between 2011 and 2022. Of these,
18,388 (17.3%) were categorised as depicting negative sentiment, and 87,852 (82.7%)
were categorised as non-negative. The categorisation of positive and neutral tweets into
non-negative was mainly due to the model's limitation in differentiating between these
types of tweets.

83
Our two-epoch training process achieved an accuracy score of 0.81 and an F1
score of 0.82 in the validation set. Specifically focussing on the classification of each
sentiment group, our final model achieved an accuracy score of 0.23 for neutral tweets,
0.44 for positive tweets and 0.83 for negative tweets in the test set. In general terms, we
can assess the accuracy metrics and F1 score as highly satisfactory (above 0.9), good
(between 0.8 and 0.9), adequate (between 0.5 and 0.8) or unsatisfactory (below 0.5).

5.4.2 Trends of non-negative perceptions over time
We identified 4,381 non-negative peaks about the Brazilian national parks between
2011 and 2022. Of these, we selected the day with the highest number of tweets per year
to identify the events that caused the greatest public interest. Our temporal analysis of
non-negative tweets about Brazilian national parks revealed the presence of several
distinct peaks (Fig. 3), with an average of 179 (83 - 377 posts/peak; SD: 93.37) nonnegative posts per annual peak. The peaks of similar events ranged from annual
celebrations of the anniversaries of certain parks, to large numbers of visitors and
government authorisations for concessions in national parks.
The highest volume of non-negative tweets (n=377) was related to a higher
visitation in Iguaçu national Park on 17th February 2015, followed by the news of the death
of an old volunteer from the Itatiaia national Park (n=284) on 28th October 2016. Of the
12 peaks of interest, 3 were related to events about the authorisation of concessions for
private companies in national parks, such as (i) Concession of the Chapada dos
Veadeiros national park on 20th December 2019 [88 tweets posted]; (ii) Government
authorisation to privatize two national parks on 10th august 2020 [83 tweets posted]; and,
(iii) concession made in the Iguaçu national park for 375 million reais on 22nd march 2022
[86 tweets posted].

Fig. 3. The daily counts related to non-negative sentiments of Twitter posts regarding
Brazilian national parks from January (2011) to December 2022. The data were obtained from
Twitter's application programming interfaces (API). The green colour represents events involving
National Parks anniversary; the purple represents the visitation in the parks; the orange represents
events related with the concessions; and, the blue colour represents others events with peaks of
public interest on Twitter.

5.4.3 Trends of negative perceptions over time
Regarding negative perceptions, we identified 3,002 peaks in Brazilian national
parks between 2011 and 2022. The peaks indicated some similar events that stimulated
the publication of negative tweets about Brazilian national parks (Fig. 4), with an average
of 171 (24 - 535 posts/peak; SD: 135.13) negative posts per annual peak. From 2011 to
2017 of the 7 events (1 per year) that caused spikes in negative tweets, 6 were related to
opinions about large-scale wildfires in national parks. The highest volume of negative

85
tweets (n=535) was related to a forest fire event in Chapada dos Guimarães national park
on 5th September 2015. However, from 2018 onwards, the events that caused the highest
negative perceptions diversified and included: (i) the oil disaster off the Brazilian coast on
2 November 2019 [108 tweets posted]; (ii) the increase in mining exploration requests in
the Amazon on 11 August 2020 [121 tweets posted]; (iii) the political attempts to reopen
a highway inside a national park on 1 June 2021 [147 tweets posted]; and (iv) the seizures
of

illegal

timber

October

2022

[70

tweets

posted].

Fig. 4. The daily counts related to negative sentiments of Twitter posts regarding Brazilian
national parks from January (2011) to December 2022. The data were obtained from Twitter's
application programming interfaces (API). The orange colours represent events involving wildfires
and the blue colours other events with peaks of public interest on Twitter.

5.4.4 Main topics associated negative perceptions

Our analysis of negative topics identified six thematic topics, based on the
clustering of the most frequent tweets within each cluster. The posts were categorized

86
according to their content across the dataset (18,388 negative tweets): (1) Wildfires (arson
and non-arson); (2) Security; (3) Regulations; (4) Wildlife roadkill; (5) Privatization; and,
(6) Lack of financial resources. The topic related to wildfires consistently appeared in all
the analyses conducted by the BERTopic model. To see the main topics performed by
BERTopic, refer to (Table 1). To gain a specific understanding of what generated negative
sentiments among the public for each park, we also explored the five most tweeted parks
between 2011-2022 and conducted a topic analysis. As the parks had different numbers
of tweets and independent subjects, they had different topic numbers. Here, they are
presented in the ranking of the most tweeted national Park. The identified topics for each
park (Fig. 5) were as follows: Iguaçu national park: (1) Regulations; (2) Political flexibility;
(3) Wildlife roadkill. Tijuca national park: (1) Security; (2) Expropriation. Itatiaia national
park: (1) Wildfires; (2) Security. Chapada dos Veadeiros national park: (1) Wildfires; (2)
Downsizing. Lençóis Maranhenses national park: (1) Downsizing; (2) Regulations; (3)
Political flexibility.

Fig. 5. A bar chart representing the dominant negative topics per park. Each topic is identified
using BERTtopic analysis and then subjectively classified based on the model output. The parks
included are the five most tweeted Brazilian Parks (2011-2022).

87
Table 1: Topics identified by the BERTopic models during the analysis of the negative tweet dataset. Models 1 to 4 were performed
with the nr_topics to auto and different min cluster sizes for the HDBSCAN parameter model.
N of models
performed
using
BERTopic

Topics

#N of
tweets

Topic named by BERTopic model

Percentage

Categorised topics

2929

1_next_christ_national forest_to do.

15.93%

Wildfires

887

2_criminals fire_arson.

4.82%

Wildfires

508

3_canastra_50_50 firefighters_flames ap_fire

2.76%

Lack of financial resources

338

4_cipó minas_gerais fires_minas gerais_wildfires

1.84%

Wildfires

336

5_df_while doing_bite_was walking

1.83%

Security

301

6_dead inside_deer found_found dead

1.64%

Wildlife roadkill

296

7_photo park_publish photo_finished publishing

1.61%

Photo publishing*

254

8_iguaçu police_environmental police_seizes

1.38%

Regulations

12539

Others and Outliers

68.19%

2896

1_christ_station_smoke_points

15.75%

Wildfires

913

2_criminals fire_arson.

4.97%

Wildfires

416

3_flames ap_fire ap_50 firefighters_ap

2.26%

Lack of financial resources

394

4_woman raped_veadeiros park_trail_sp

2.14%

Security

313

5_alerts_km park_inpe_spacial

1.70%

Wildfires

302

6_photo park_publish photo_finished publishing

1.64%

Photo publishing*

296

7_dead inside_deer found_found dead

1.61%

Wildlife roadkill

8
-

276

8_environmental police_remove cattle

1.50%

Regulations

12582

Others and Outliers

68.43%

-1

2960

1_station_smoke_christ_larger

16.10%

Wildfires

912

2_criminals fire_arson_wildfires

4.96%

Wildfires

326

3_protest_Iguaçu decision_taxis_federal prohibit

1.77%

Regulations

310

4_dead inside_deer found_found dead

1.69%

Wildlife roadkill

293

5_ iguaçu police_environmental police

1.59%

Regulations

289

6_destroys area_hit 12_12 thousand_main

1.57%

Wildfires

277

hotspots

1.51%

Photo publishing*

8
-

238

7_ photo park_publish photo_finished

1.29%

Privatization

12783

8_privatization_privatization park_privatize

69.52%

Others and Outliers
4

2896

1_christ_station_smoke_points

15.75%

Wildfires

913

2_criminals fire_arson_wildfires

4.97%

Wildfires

416

3_flames ap_fire ap_50 firefighters_ap

2.26%

Lack of financial resources

394

4_woman raped_veadeiros park_trail_sp

2.14%

Security

313

5_alerts_km park_inpe_spacial

1.70%

Wildfires

302

6_ photo park_publish photo_finished publishing

1.64%

Photo publishing*

296

7_dead inside_deer found_found dead

1.61%

Wildlife roadkill

276

8_environmental police_remove cattle

1.50%

Regulations

12582

Others and Outliers

68.43%

5.5 Discussion
This study provides the first assessment of perceptions related to Brazilian national
parks, employing natural language processing techniques to analyse textual content
shared on the social media platform Twitter.
5.5.1 Non-negative perceptions of Brazilian National parks
Our results revealed that users of the social media Twitter expressed a nonnegative (neutral or positive) sentiment towards Brazilian national parks, which is in line
with the results of previous studies by Hausmann et al. (2018) and Cao et al. (2022). The
non-negative sentiment results are probably driven by many factors, including that the
Brazilian national park system has experienced a number of celebratory milestones in
recent years. The anniversary celebrations of the creation of the Itatiaia and Iguaçu
national parks, as well as high visitor rates on commemorative dates in the Iguaçu national
park, generated great public interest among Twitter users. An interesting result is almost
50% of the peaks of public interest related to Iguaçu national park (Fig.3), corroborating
recent studies that have shown that it is the most mentioned Brazilian park on the national
and global internet (Correia et al., 2018) and that it is also the most tweeted protected
area (among all Brazilian categories) in the period from 2011 to 2020 (Souza et al. 2023).
This may be due to its exceptional natural attributes and high annual number of visitors
(ICMBio, 2022). In addition, Iguaçu national park was the first Brazilian park to be granted
to a private company, leading to higher investment and greater media exposure.
However, in addition to the clearly positive findings, we also identified events that
provoke public interest, but which are problematic for the classification of sentiments. The
first of these is the news related to the death of a volunteer in Itatiaia national park. Such
an event can evoke feelings of sadness and grief, but also be written about in positive
terms celebrating the importance of the park to the deceased person. Another event that
provoked negative sentiments was the attempt to reopen a road in Iguaçu national park
and the authorisation of concessions for national parks to private initiatives, revealing
possible conflicts between development aspirations and conservation needs.

90
In situations where the political and social context is mixed analysing sentiment
becomes more challenging (Avanço and Nunes, 2014). As mentioned (section 3.1)
previously, differentiating between positive and neutral sentiments was exceedingly
challenging. Generally, news stories are evaluated by algorithms as neutral, because they
don't contain words that can be classified as positive and/or negative. For example,
national park concessions to private initiatives are widely reported by the media, often
generating debate about environmental impacts and nature conservation.
Analysing sentiment in such a diverse context requires a more sophisticated
approach, capable of capturing the duality of emotions present in different topics, offering
a more balanced view of public reactions. Therefore, given the weaker accuracy in the
classification of non-negative tweets (0.23 for neutral tweets; 0.44 for positive tweets), and
considering that negative tweets express negative feelings more strongly, often related to
issues and conflicts perceived as directly relevant to improving biodiversity conservation
and the management of these areas, we chose to discuss the results related to negative
feelings in more detail.

5.5.2 Negative perceptions of Brazilian National parks
As mentioned by Fink et al., (2020), negative events tend to elicit stronger reactions
from the public. However, it is worth noting that, in general, negative events associated
with PAs not only tend to trigger an immediate response, as in the case of reaction on
social media platforms, but also tend to have lasting negative impacts on attitudes towards
these PAs (Bragagnolo et al. 2015). For instance, displacements of people from protected
lands and conflicts between natural resource users and PA managers have long-term
negative impacts on individuals, even when these actions result in positive conservation
outcomes (Brumatti and Rozendo, 2021; Maciel, 2015; Rossi et al., 2016). From 2011 to
2017, wildfires were the main driver of negative sentiments about the national parks,
representing on average 22.51% of the negative perceptions of all four models (Table 1).
The media played a critical role in disseminating these events, with a substantial
proportion of the tweets originating from news agencies. It is well-known that the media
(traditional or online) exert influence over people's perception regarding various

91
environmental and political issues (Shah et al., 2007). When combined with popular social
media discussion platforms, these agencies become ideal means to promote public
engagement and can also be effective in shaping public perceptions and mobilising realworld actions (Almeida et al., 2022; Stanley, 2020). This further supports the observation
that Twitter data can be employed to measure the public response to environmental risks
and hazards, such as wildfires (Shook and Turner, 2016).
Our results revealed that the majority of news regarding fires were related to
national parks of the Cerrado region and were posted during the dry season (Fig.4). Even
in the Brazilian Cerrado, an ecosystem in which fire plays a fundamental role in
evolutionary terms and the maintenance of crucial ecological processes, fire is often
related to human activities (including climate change) and driven by agricultural practices
generating a strong external pressure derived from land use changes in the national park’s
vicinities (Pivello et al., 2021). In addition, high-impact wildfire events are subjects of great
media interest, and the intensive dissemination of information about the dangers during a
short period of imminent disaster threat can sensitise people to the impending event (Perry
et al., 1982). Thus, the way news is conveyed, taking into account the factual aspects,
plays an important role in raising awareness among both the public and policymakers
regarding this issue.
Over time, our findings suggest that while wildfires were the main factor of negative
sentiment towards national parks from 2011 to 2017, other issues related to environmental
and political concerns have emerged as important factors in recent years. These include
oil spills on the Brazilian coast (Almeida, et al 2022), mining requests in protected areas
(Siqueira-Gay et. al 2022), political attempts to reopen a road within a national park
(Prasniewski et al. 2020), and the seizure of illegal timber in an Amazonian national park
(Alencar et al. 2022). Brazilian environmental policy has been subject to criticism and
controversy in recent years (Dobrovolski et. al. 2018), especially from 2019 onwards,
when the federal government began to relax environmental laws and reduce surveillance
and protection of protected areas (Barbosa et al., 2021; Fearnside, 2019). This has
generated high levels of concern and reactions from civil society, academia, and national
and international organisations, as the government considered environmental restrictions
and procedures as obstacles to progress (Abessa et al., 2019). This may help explain the

92
peaks of negative tweets related to Brazilian national parks in the analysed period. It is
worth noting that wildfires have not ceased to occur; however, the emphasis on
environmental disregard of the federal government of 2019 drew attention to the decisions
being made in this area. For example, the oil spill in northeast Brazil led to many Instagram
users expressing despair over the government's inaction and lack of response (Almeida
et al. 2022).
In addition to our temporal analysis, our aim was also to identify the most discussed
topics over the past 12 years. Our topic modelling analysis identified r topics such as
wildfires and wildlife roadkill, which are typically isolated events generating heightened
engagement. Topic analysis is a powerful tool for attaining a comprehensive and in-depth
understanding of the content within a dataset (Grootendorst, 2022). Notably, the presence
of cattle grazing and the insufficiency of financial resources (Table 1) emerged as
intriguing findings in the analysis, highlighting conflicts stemming from the agricultural
sector's push to introduce cattle into protected areas, an activity prohibited in national
parks, coupled with the challenges in fostering collaboration between the management
body of these areas and cattle breeders to establish inclusive and participatory
management agreements (Borges et al., 2014). It also draws attention to the daunting
challenge of resource scarcity (Gerhardinger et al., 2011) in protected area management.
Brazilian national parks, on average, have just one staff member per 11,000 hectares
(Instituto Semeia, 2021). This broader perspective also underscores the potency of social
media and sentiment analysis in monitoring trends, patterns, and reactions pertaining to
management and conservation-related matters (Fink et al., 2020; Soriano-Redondo et al.,
2017) and its ability to provide fine scale context dependent information.
In examining the management of the most tweeted parks, we identified specific
topics that yielded distinctive results for each park. Negative issues were identified that
converge between the parks, and distinct topics were also identified for each park (Fig.
5), probably due to their unique characteristics and management factors. Among the
topics uncovered, the most prominent intersection revolved around regulations. Several
regulations were found to have adverse effects (Refer to Table 2 in the supplementary
material for details), from the prohibition of taxis in Iguaçu national park to the prohibition
of quad bikes in Lençóis Maranhenses national park. It is widely recognised that

93
management plan restrictions can promote negative perceptions regarding the use of
space and can lead to a disconnection between society and protected areas, thus
undermining appreciation of the importance of nature (Hausmann et al., 2020). On the
other hand, it is also important to acknowledge that practices such as quad biking can
both generate value for participants while diminishing value for other park users (e.g.
birdwatchers). In addition to regulations, problems related to expropriation in the Tijuca
national park, the possibility of reducing the size of the Lençóis Maranhenses national
park, and political flexibilisation for the creation of a new category of protected area
(Estrada-Parque) in the Iguaçu national park are external pressures that can be related
to PADDD events. Bernard and colleagues (2014) identified 41 PADDD events that
occurred in Brazil from 1979 to 2012 as a result of chronic deficiencies in financial,
personnel and enforcement resources. Government agencies often implement PADDD
without consulting civil society, jeopardising the integrity of Protected Areas. These areas
exemplify common public goods that require a set of robust governance practices that
take into account the complexity of socio-environmental systems, transcend conflicts of
interest regarding ecosystems and safeguard the legitimacy of decision-making
processes and spaces. In this sense, (Macedo and Medeiros, 2018) have already
proposed that participatory incentives (in a non-top-down approach) and knowledge
incentives serve as the main drivers of cooperation and effectiveness in protected areas.
The adoption of transparent and communicative measures between individuals and
institutions is imperative to increase society's interest in the management of conservation
territories and promote the inclusion of all social actors in decision-making spaces (Souza
et al., 2022).

5.5.3 Potential, limitations and future research
Overall, sentiment analysis proved to be a valuable methodology for discerning
conflicts, such as cases of expropriation, cattle grazing, disagreements linked to
restrictions, and safety concerns within national parks (Fig. 5), thus furnishing a potent
analytical instrument for protected area management. Monitoring public sentiment through
social media could be a way to: (i) monitor events that cause more public interest and
negative engagement, enabling rapid action by protected area management to crises and

94
adverse events; (ii) identify gaps through topic analysis and improve strategies to
maximise positive results and minimise negative impacts, such as creating fire brigades
and environmental education and interpretation strategies on fire in protected areas; (iii)
understand the public's perception of and response to any park use rules, promoting more
effective communication, addressing concerns, clarifying misunderstandings and
fostering trust and collaboration between society and protected area management. Still,
the application of sentiment analysis within this context introduces several intertwined
challenges and limitations that merit careful consideration. These challenges encompass
issues concerning data quality (Mozetič et al., 2016), the presence of subjectivity and irony
within textual content (Ravi and Ravi, 2015), along with concerns relating to
representativeness (Di Minin et al., 2015) and ethical considerations tied to data collection
(for a comprehensive discussion, refer to (Di Minin et al., 2021). Moreover, during our
analysis, we noted that the datasets employed to train the model, encompassing both the
categorised Twitter dataset and the dataset of opinions about online products,
inadequately distinguished between positive and neutral sentiments, resulting in their
aggregated categorization as non-negative. It is pivotal to emphasise that the efficacy of
classification models is more reliant on the quality, representativeness, and extent of
training data than on the particular model type employed (Di Minin et al., 2015; Mozetič et
al., 2016). Beyond this limitation, we also observed that sentiments expressed by Twitter
users concerning Brazilian national parks, even when perceived as neutral, frequently
mirrored a favourable perception and general appreciation of experiences within protected
areas. However, it is important to underscore that the dataset employed (B2WReviews01, 2018; Corpus Buscapé, 2013) exhibited satisfactory performance in
identifying tweets containing negative sentiments, which are the primary focus of this
study.
Despite these challenges and limitations, sentiment analysis can be a valuable tool
in the field of conservation management, if approached carefully, and has the potential to
contribute to the sustainable management and effective governance of protected areas.
By extending the use of sentiment analysis in conservation to the Portuguese language,
our research outlines new avenues for this research domain that can focus on developing
customised approaches for languages other than English. Given the widespread use of

95
the Portuguese language, which is spoken by more than 200 million people worldwide
(Instituto Camões, 2022), our study offers an avenue for gaining insights into the opinions
and attitudes of the population by analysing the sentiments expressed in Portuguese
about the protected areas present in Portuguese-speaking countries.
For future work, we recommend the use of a pre-categorised sentiment database
focused on environmental themes, which can help increase the accuracy of machine
learning models and obtain better results in the classification of positive and neutral
sentiments. In addition, we propose incorporating field studies into future research, thus
bringing together and validating online results with personal perceptions of protected
areas. This multi-faceted approach would produce a more holistic understanding of
human-nature interactions in protected areas.
REFERENCES
Abessa, D., Famá, A., Buruaem, L., 2019. The systematic dismantling of Brazilian
environmental laws risks losses on all fronts. Nat Ecol Evol. https://doi.org/10.1038/
s41559-019-0855-9.
Agyeman, Y.B., Aboagye, O.K., Ashie, E., 2019. Visitor satisfaction at Kakum National
Park in Ghana. Tour. Recreat. Res. 44, 178–189. https://doi.org/10.1080/
02508281.2019.1566048.
Ahani, A., Nilashi, M., Yadegaridehkordi, E., Sanzogni, L., Tarik, A.R., Knox, K., Samad,
S., Ibrahim, O., 2019. Revealing customers’ satisfaction and preferences through online
review analysis: the case of Canary Islands hotels. J. Retail. Consum. Serv. 51, 331–
343. https://doi.org/10.1016/j. jretconser.2019.06.014.
Alencar, A., Silvestrini, R., Gomes, J., Savian, G., 2022. Amazônia em chamas: o novo
e alarmante patamar do desmatamento na Amazônia. Nota Técnica 9.
Almeida, J.A.G.R., Guedes-Santos, J., Vieira, F.A.S., Azevedo, A.K., Souza, C.N.,
Pinheiro, B.R., Correia, R.A., Malhado, A.C.M., Ladle, R.J., 2022. Public awareness and
engagement in relation to the coastal oil spill in Northeast Brazil. An. Acad. Bras.Cienc.
94, 1–10. https://doi.org/10.37002/biobrasil.v12i2. 2177.
Avanço L.M. and Nunes, M. D. G. V. Lexicon-Based Sentiment Analysis for Reviews of
Products in Brazilian Portuguese. 2014. Brazilian conference on intelligent systems, São
Paulo, Brazil, 2014, pp. 277–28. doi:https://doi.org/10.1109/BRACIS.2014.57.
B2W-Reviews01., 2021. Open corpus of product reviews. GitHub. https://github.com/
b2wdigital/b2w-reviews01.

96
Barbosa, L.G., Alves, M.A.S., Grelle, C.E.V., 2021. Actions against sustainability:
dismantling of the environmental policies in Brazil. Land Use Policy 104. https://
doi.org/10.1016/j.landusepol.2021.105384.
Becken, S., Stantic, B., Chen, J., Alaei, A.R., Connolly, R.M., 2017. Monitoring the
environment and human sentiment on the great barrier reef: assessing the potential of
collective sensing. J. Environ. Manage. 203, 87–97. https://doi.org/10.1016/
j.jenvman.2017.07.007.
Bernard, E., Penna, L.A.O., Araújo, E., 2014. Downgrading, downsizing, degazettement,
and reclassification of protected areas in Brazil. Conserv. Biol. 28, 939–950. https://
doi.org/10.1111/cobi. 12298.
Bhatt, P., Pickering, C.M., 2021. Public perceptions about Nepalese National Parks: A
global twitter discourse analysis. Soc. Nat. Resour. 34, 683–700. https://doi.org/
10.1080/08941920.2021.1876193.
Birjali, M., Kasri, M., Beni-Hssane, A., 2021. A comprehensive survey on sentiment
analysis: approaches, challenges and trends. Knowl Based Syst 226. https://doi.org/
10.1016/j.knosys.2021.107134.
Borges, S.L., Eloy, L., Ludewigs, T., 2014. O gado que circulava: desafios da gestão
participativa e impactos da proibição do uso do fogo aos criadores de gado de solta da
Reserva de Desenvolvimento Sustentável Veredas do Acari. Biodiversidade Brasileira. 4
(1), 130–156.
Bragagnolo, C., Costa Gamarra, N., Malhado, Claudia Mendes, A., James Ladle, R.,
2016. Proposta Metodológica para Padronização dos Estudos de Atitudes em
Comunidades Adjacentes às Unidades de Conservação de Proteção Integral no Brasil.
Biodiversidade Brasileira 6 (1), 190–208.
Bragagnolo, C., Correia, R.A., Gamarra, N.C., Lessa, T., Jepson, P., Malhado, A.C.M.,
Ladle, R.J., 2021. Uncovering assets in Brazilian national parks. J. Environ. Manage.
287. https://doi.org/10.1016/j.jenvman. 2021.112289.
BRASIL, 2000. Lei No 9.985, de 18 de julho de 2000. Institui o Sistema Nacional de
Unidades de Conservação da Natureza e dá outras providências. (Law.)
Brumatti, P.N.M., Rozendo, C., 2021. Parques Nacionais, turismo e governança.
Revista Brasileira de Pesquisa em Turismo 15, 2119. https://doi.org/10.7784/
rbtur.v15i3.2119.
Cao, H., Wang, M., Su, S., Kang, M., 2022. Explicit quantification of coastal cultural
ecosystem services: A novel approach based on the content and sentimental analysis of
social media. Ecol. Indic. 137. https://doi.org/10.1016/j.ecolind.2022.108756.
Capellaro, L., 2021. Análise de polaridade e de tópicos em tweets no domínio da
política no Brasil. São Carlos, São Paulo.

97
Ceron, A., Curini, L., Iacus, S.M., Porro, G., 2014. Every tweet counts? How sentiment
analysis of social media can improve our knowledge of citizens’ political preferences
with an application to Italy and France. New Media Soc. 16, 340–358. https://
doi.org/10.1177/1461444813480466.
CNUC, 2023. Painel de Unidades de Conservação. https://cnuc.mma.gov.br/powerbi
(accessed 06 November 2023).
Corpus Buscapé, 2013. Portuguese product reviews https://drive.google.com/file/d/
1IZJuvt1 uxQ4oPGAvGQQxQ_h_ZiV-Be72/view.
Correia, R.A., Jepson, P., Malhado, A.C.M., Ladle, R.J., 2018. Culturomic assessment
of Brazilian protected areas: exploring a novel index of protected area visibility. Ecol.
Indic. 85, 165–171. https://doi.org/10.1016/j.ecolind.2017.10.033.
Correia, R.A., Ladle, R., Jarić, I., Malhado, A.C.M., Mittermeier, J.C., Roll, U., SorianoRedondo, A., Veríssimo, D., Fink, C., Hausmann, A., Guedes-Santos, J., Vardi, R., Di
Minin, E., 2021. Digital data sources and methods for conservation culturomics.
Conserv. Biol. 35, 398–411. https://doi.org/10.1111/cobi.13706.
De Haan, F.J., Ferguson, B.C., Adamowicz, R.C., Johnstone, P., Brown, R.R., Wong,
T.H.F., 2014. The needs of society: A new understanding of transitions, sustainability
and liveability. Technol Forecast Soc Change 85, 121–132. https://doi.org/10.1016/
j.techfore.2013.09.005.
Di Minin, E., Tenkanen, H., Toivonen, T., 2015. Prospects and challenges for social
media data in conservation science. Front. Environ. Sci. 3, 1–6. https://doi.org/10.3389/
fenvs.2015.00063.
Di Minin, E., Fink, C., Hausmann, A., Kremer, J., Kulkarni, R., 2021. How to address
data privacy concerns when using social media data in conservation science. Conserv.
Biol. 35, 437–446. https://doi.org/10.1111/cobi.13708.
Dobrovolski, R., Loyola, R., Rattis, L., Gouveia, S.F., Cardoso, D., Santos-Silva, R.,
Gonçalves-Souza, D., Bini, L.M., Diniz-Filho, J.A.F., 2018. Science and democracy must
orientate Brazil’s path to sustainability. Perspectives in Ecology and Conservation 16 (3),
121–124.
Drijfhout, M., Kendal, D., Vohl, D., Green, P.T., 2016. Sentiment analysis: ready for
conservation. Front. Ecol. Environ. 14, 525–526. https://doi.org/10.1002/fee.1435.
Dudley, N., 2008. Guidelines for Applying Protected Area Management Categories.
IUCN, IUCN, Gland, Switzerland. https://doi.org/10.1103/PhysRevB.38.10724.
Egger, R., Yu, J., 2022. A topic modeling comparison between LDA, NMF, Top2Vec,
and BERTopic to demystify twitter posts. Front. Sociol. 7. https://doi.org/10.3389/
fsoc.2022.886498.

98
Fearnside, P.M., 2019. Setbacks under president Bolsonaro: A challenge to
sustainability in the Amazon. Sustentabilidade International Science Journal 1, 38–52.
Ferreira, J., et al., 2014. Brazil’s environmental leadership at risk. Science 346, 706–
707. https://doi.org/10.1126/science.1260194.
Fink, C., Hausmann, A., Di Minin, E., 2020. Online sentiment towards iconic species.
Biol. Conserv. 241. https://doi.org/10.1016/j.biocon.2019.108289.
Fredrickson, B.L., 2001. The role of positive emotions in positive psychology: the
broaden- and-build theory of positive emotions. Am. Psychol. 56, 218–226.
https://doi.org/ 10.1037/0003-066X.56.3.218.
Gamarra, N.C., Correia, R.A., Bragagnolo, C., Campos-Silva, J.V., Jepson, P.R., Ladle,
R.J., Mendes Malhado, A.C., 2019. Are protected areas undervalued? An asset-based
analysis of Brazilian protected area management plans. J. Environ. Manage. 249,
109347. https://doi.org/10.1016/j.jenvman.2019.109 347.
Gerhardinger, L.C., Godoy, E.A.S., Jones, P.J.S., Sales, G., Ferreira, B.P., 2011. Marine
protected dramas: the flaws of the Brazilian national system of marine protected areas.
Environ. Manag. 47, 630–643. https://doi.org/10.1007/s00267-010-9554-7.
Griggs, N., Lacey, G.T., 2022. Barriers and limitations to national park visitation by
millennials: perceptions from second-generation Australians. Annals of Tourism
Research Empirical Insights 3. https://doi.org/10.1016/j.annale.2022.100074.
Grootendorst, M., 2022. BERTopic: Neural Topic Modeling with a Class-Based TF-IDF
Procedure.
Hausmann, A., Toivonen, T., Slotow, R., Tenkanen, H., Moilanen, A., Heikinheimo, V.,
Di Minin, E., 2018. Social media data can be used to understand Tourists’ preferences
for nature-based experiences in protected areas. Conserv. Lett. https://doi.org/10.1111/
conl.12343.
Guedes-Santos, J., Correia, R.A., Jepson, P., Ladle, R.J., 2021. Evaluating public
interest in protected areas using Wikipedia page views. J Nat Conserv 63.
https://doi.org/ 10.1016/j.jnc.2021.126040.
Hausmann, A., Toivonen, T., Fink, C., Heikinheimo, V., Kulkarni, R., Tenkanen, H., Di
Minin, E., 2020. Understanding sentiment of national park visitors from social media
data. People and Nature pan 3.10130. https://doi.org/10.1002/pan3.10130.
Hockings, M., Stolton, S., Leverington, F., 2006. Evaluating Effectiveness : A Framework
for Assessing Management Effectiveness of Protected Areas, 2nd Edition, Evaluating
Effectiveness : A Framework for Assessing Management Effectiveness of Protected
Areas, 2nd Edition. https://doi.org/10.2305/iucn.ch.2006.pag. 14
Hoeffel, J.L., Fadini, A.A.B., Machado, M.K., Reis, J.C., 2008. Trajetórias do JaguaryUnidades de conservação, percepção ambiental e turismo: Um estudo na apa do

99
sistema Cantareira, São Paulo. Ambiente e Sociedade 11, 131–148. https://doi.org/
10.1590/s1414-753x2008000100010.
ICMBio, 2022. Unidades de conservação federais recebem mais de 21 milhões de
visitas em 2022. https://www.gov.br/icmbio/pt-br/assuntos/noticias/ultimas-noticias/
unidades-de-conservacao-federais-recebem-mais-de-21-milhoes-de-visitas-em-2022
(accessed 20 November 2023).
Instituto Camões, 2022. Dados sobre a Língua Portuguesa. https://www.institutocamoes.pt/images/ pdf_noticias/Dados_sobre_a_l%C3%ADngua_portuguesa_de_
2022.pdf (accessed 20 April 2023).
Instituto Semeia, 2021. Diagnóstico do uso público em parques brasileiros: a
perspectiva da gestão.
Instituto Semeia, 2022. Parques do Brasil: percepções da População 2022.
Jepson, P.R., Caldecott, B., Schmitt, S.F., Carvalho, S.H.C., Correia, R.A., Gamarra, N.,
Bragagnolo, C., Malhado, A.C.M., Ladle, R.J., 2017. Protected area asset stewardship.
Biol. Conserv. 212, 183–190. https://doi.org/10.1016/j.biocon.2017.03.032.
Kaity, M., Balakrishnan, V., 2020. Sentiment lexicons and non-English languages: a
survey. Knowl. Inf. Syst. 62 (12), 4445–4480.
Kim, H., Shoji, Y., Mameno, K., Kubo, T., Aikoh, T., 2023. Changes in visits to green
spaces due to the COVID-19 pandemic: focusing on the proportion of repeat visitors and
the distances between green spaces and visitors’ places of residences. Urban For.
Urban Green. 80, 127828.
Ladle, Richard J., Souza, Carolina N., Correia, R., 2021. Culturomics for (not against!)
protected areas in. Biol. Conserv. 256, 109197. https://doi.org/10.1016/
j.biocon.2021.109015.
Lemberg, D., 2010. Environmental Perception. In: Warf, B., Encyclopedia of Geography.
Sage.
Liu, B., 2012. Sentiment Analysis and Opinion Mining. Morgan & Claypool.
Lubbe, B.A., du Preez, E.A., Douglas, A., Fairer-Wessels, F. The impact of rhino
poaching on tourist experiences and future visitation to National Parks in South Africa.
Current Issues in Tourism. https://doi.org/10.1080/13683500.2017.1343807.
Macedo, H.S., Medeiros, R.P., 2018. Rethinking governance in a Brazilian multiple-use
marine protected area. Mar. Policy 0–1. https://doi.org/10.1016/ j.marpol.2018.08.019.
Maciel, G.G., 2015. Mercantilização da cidade do Rio de Janeiro e suas implicações na
gestão de unidades de conservação: um estudo sobre a concessão do Setor Paineras/

100
Corcovado (Parque Nacional da Tijuca-RJ) e os efeitos sobre os moradores das favelas
do Cerro Corá e do Guararapes. Pontifícia Universidade Católica do Rio de Janeiro, Rio
de Janeiro.
Maretti, C.C., Catapan, M., Abreu, M.J., Oliveira, J.E.D., 2012. Áreas protegidas:
definições, tipos e conjuntos – reflexões conceituais e diretrizes para a gestão.
Mistry, Jayalaxshmi, Bizerril, Marcelo, 2011. Por Que é Importante Entender as InterRelações entre Pessoas, Fogo e Áreas Protegidas? Biodiversidade Brasileira I, 40–49.
https://doi.org/10.37002/biobrasil.v% 25vi%25i.137.
Mozetič, I., Grčar, M., Smailović, J., 2016. Multilingual twitter sentiment classification:
the role of human annotators. PloS One 11. https://doi.org/10.1371/
journal.pone.0155036.
Nabout, J.C., Tessarolo, G., Pinheiro, G.H.B., Marquez, L.A.M., de Carvalho, R.A.,
2022. Unraveling the paths of water as aquatic cultural services for the ecotourism in
Brazilian protected areas. Global Ecology and Conservation 33, e01958.
Otsuka, R., Yamakoshi, G., Id, R.O., Yamakoshi, G., 2020. Analyzing the popularity of
YouTube videos that violate mountain gorilla tourism regulations. PloS One 15, 1–20.
https://doi.org/10.1371/journal. pone.0232085.
Papworth, S.K., Nghiem, T.P.L., Chimalakonda, D., Posa, M.R.C., Wijedasa, L.S.,
Bickford, D., Carrasco, L.R., 2015. Quantifying the role of online news in linking
conservation research to Facebook and twitter. Conserv. Biol. 29, 825–833.
https://doi.org/ 10.1111/cobi.12455.
Pereira, D.A., 2021. A survey of sentiment analysis in the Portuguese language. Artif.
Intell. Rev. 54, 1087–1115. https://doi.org/10.1007/s10462-020-09870-1.
Perry, R.W., Lindell, M.K., Greene, M.R., 1982. Threat perception and public response
to volcano Hazard. J. Soc. Psychol. 116 (2), 199–204.
https://doi.org/10.1080/00224545.1982.9922771.
Pivello, V.R., 2006. Fire management for biological conservation in the Brazilian
cerrado. In: Mistry, J., Berardi, A. (Eds.), Savannas and Dry Forests: Linking People with
Nature. Ashgate Publications, pp. 129–154.
Portuguese tweets for Sentiment Analysis | Kaggle, 2018. GitHub.
https://www.kaggle.com/ datasets/augustop/portuguese-tweets-for-sentiment-analysis/
download?datasetVersionNumber=2.
Prasniewski, V.M., Szinwelski, N., Sobral-Souza, T., Kuczach, A.M., Brocardo, C.R.,
Sperber, C.F., Fearnside, P.M., 2020. Parks under attack: Brazil’s Iguaçu National Park
illustrates a global threat to biodiversity. Ambio. https://doi.org/10.1007/ s13280-02001353-5.

101
Ravi, K., Ravi, V., 2015. A survey on opinion mining and sentiment analysis: tasks,
approaches and applications. Knowl Based Syst 89, 14–46. https://doi.org/10.1016/
j.knosys.2015.06.015.
Rossi, S.D., Byrne, J.A., Pickering, C.M., Reser, J., 2015. “Seeing red” in national parks:
how visitors’ values affect perceptions and park experiences. Geoforum 66, 41–52.
https://doi.org/10.1016/j.geoforum.2015. 09.009.
Rossi, S.D., Pickering, C.M., Byrne, J.A., 2016. Not in our park! Local community
perceptions of recreational activities in peri-urban national parks. Aust. J. Environ.
Manag. 23, 245–264. https://doi.org/10.1080/14486563.2015.1132397.
Rylands, A.B., Brandon, K., 2005. Brazilian protected areas. Conserv. Biol.
doi:10.1111/j.1523–1739.2005.00711.x.
Serrano-Guerrero, J., Olivas, J.A., Romero, F.P., Herrera-Viedma, E., 2015. Sentiment
analysis: A review and comparative analysis of web services. Inf Sci (N Y) 311, 18–38.
https://doi.org/10.1016/j.ins.2015.03. 040.
Shah, D.V., McLeod, D.M., Kim, E., Lee, Sun Young, Gotlieb, M.R., Ho, S.S., Breivik, H.,
2007. Political consumerism: how communication and consumption orientations drive
lifestyle politics. Annals of the American Academy of Political and Social Science 611,
217–235. https://doi.org/10.1177/0002 716206298714.
Shook, E., Turner, V.K., 2016. The socio-environmental data explorer (SEDE): a social
media–enhanced decision support system to explore risk perception to hazard events.
Cartogr. Geogr. Inf. Sci. 43, 427–441. https://doi.org/10.1080/ 15230406.2015.1131627.
Sievert, C., 2020. Interactive Web-Based Data Visualization with R, Plotly, and Shiny.
https://doi.org/10.1201/9780429447273.
Silva, J.M.C. da, Dias, T.C.A. de C., Cunha, A.C. da, Cunha, H.F.A., 2021. Funding
deficits of protected areas in Brazil. Land Use Policy 100. doi:https://doi.org/10.1016/
j.landusepol.2020.104926 .
Siqueira-Gay, J., Metzger, J.P., Sánchez, L.E., Sonter, L.J., 2022. Strategic planning to
mitigate mining impacts on protected areas in the Brazilian Amazon. Nat Sustain 5, 853–
860. https://doi.org/10.1038/s41893-022-00921-9.
Soares-Filho, B., Moutinho, P., Nepstad, D., Anderson, A., Rodrigues, H., Garcia, R.,
Dietzsch, L., Merry, F., Bowman, M., Hissa, L., Silvestrini, R., Maretti, C., 2010. Role of
Brazilian Amazon protected areas in climate change mitigation. Proc. Natl. Acad. Sci. U.
S. A. 107, 10821–10826. https://doi.org/10.1073/pnas. 0913048107.
Soriano-Redondo, A., Bearhop, S., Lock, L., Votier, S.C., Hilton, G.M., 2017. Internetbased monitoring of public perception of conservation. Biol. Conserv. 206, 304–309.
https://doi.org/10.1016/j.biocon. 2016.11.031.

102
Souza, F., Nogueira, R., Lotufo, Roberto, 2020. BERTimbau: Pretrained BERT Models
for Brazilian Portuguese, in: 9th Brazilian Conference. pp. 403–417. doi:https://doi.org/
10.1007/978-3-030-61377-8
Souza, C.N., de Barros, E.L.S.F.C., Dantas, I.F.V., Bragagnolo, C., Malhado, A.C.M.,
Selva, V.F., 2022. Inclusion and governance in the managing Council of the Costa dos
Corais environmental protection area. Ambiente e Sociedade. 25.
https://doi.org/10.1590/ 1809-4422ASOC20210074R1VU2022L3AO.
Souza, C.N., Almeida, J.A.G.R, Correia R.A, Ladle R.J, Carvalho A.R, Malhado A.C.M.
2023. Assessing Brazilian protected areas through social media: insights from 10 years
of public interest and engagement. PloS One 18(10): e0293581. doi:https://
doi.org/10.1371/journal.pone.0293581.
Stanley, P., 2020. Unlikely hikers? Activism, Instagram, and the queer mobilities of fat
hikers, women hiking alone, and hikers of colour. Mobilities 15, 241–256. https://
doi.org/10.1080/17450101.2019.1696038.
Sudhir, P., Suresh, V.D., 2021. Comparative study of various approaches, applications
and classifiers for sentiment analysis. Global Transitions Proceedings 2, 205–211.
https:// doi.org/10.1016/j.gltp.2021.08.004.
Team, R.C, 2017. R: A Language and Environment for Statistical Computing. Tenkanen,
H., Di Minin, E., Heikinheimo, V., Hausmann, A., Herbst, M., Kajala, L.,
Toivonen, T., 2017. Instagram, Flickr, or twitter: assessing the usability of social media
data for visitor monitoring in protected areas. Sci. Rep. 7. https://doi.org/
10.1038/s41598-017-18007-4.
Toivonen, T., Heikinheimo, V., Fink, C., Hausmann, A., Hiippala, T., Järv, O., Tenkanen,
H., Di Minin, E., 2019. Social media data for conservation science: A methodological
overview. Biol Conserv 233, 298–315. https://doi.org/10.1016/ j.biocon.2019.01.023.
Vale, M.M., Berenguer, E., Argollo de Menezes, M., Viveiros de Castro, E.B., Pugliese
de Siqueira, L., Portela, R.Q., 2021. The COVID-19 pandemic as an opportunity to
weaken environmental protection in Brazil biol. Conserv 255, 108994.
Watson, J.E.M., Dudley, N., Segan, D.B., Hockings, M., 2014. The performance and
potential of protected areas. Nature 515, 67–73. https://doi.org/10.1038/ nature13947.
Wickham, H., 2008. Elegant Graphics for Data Analysis: ggplot2. (Applied Spatial Data
Analysis with R).
Wickham, H.; François, R.; Henry, L.; Müller, K., 2021. Dplyr: A grammar of data
manipulation.
Wilkins, L., 1985. Television and newspaper coverage of blizzard: is the message
helplessness? Newsp. Res. J. 51–65.

103
6 CONSIDERAÇÕES FINAIS
Este estudo oferece uma contribuição significativa ao crescente campo da análise
de métricas culturômicas para monitorar as interações homem-natureza em grandes
escalas. A análise de dados das redes sociais, demonstrou ser uma poderosa ferramenta
capaz de promover uma abordagem mais abrangente e participativa para a gestão das
áreas protegidas (APs) brasileiras.
No primeiro artigo, a análise de 10 anos de discurso no Twitter sobre as APs
destacou o poderoso potencial das redes sociais para a compreensão das percepções
sobre essas áreas. O estudo realçou a necessidade de aprimorar a comunicação oficial
sobre as APs, destacando a importância de fornecer informações básicas, como o
significado dessas áreas, esclarecimento das diferenças entre áreas protegidas e
parques urbanos, e a relevância dos regulamentos específicos de cada AP.
Adicionalmente, o estudo destacou que os usuários das redes sociais demonstram maior
envolvimento com publicações relacionadas à conflitos. O monitoramento das redes
sociais pode auxiliar os gestores na compreensão das percepções e interesses do
público que utiliza estas plataformas online, possibilitando a identificação de lacunas e a
antecipação de possíveis conflitos para aprimorar estratégias visando maximizar os
resultados positivos das ações de conservação (Capítulo 1, Fig. 7A e 7B).
No segundo artigo, a análise de sentimentos emergiu como uma ferramenta
valiosa para a gestão de APs, fornecendo insights sobre as percepções negativas
manifestadas pelos usuários do Twitter com relação aos parques nacionais brasileiros.
Foram identificadas áreas críticas abrangendo temas como incêndios nos parques
nacionais, segurança dentro do parque, mortalidade da vida selvagem por
atropelamentos, regramentos, privatizações (concessões) e a escassez de recursos
financeiros para a otimização da gestão dessas áreas (Capítulo 2, Tabela 1). Além disso,
diminuímos a escala espacial para compreender o desempenho da ferramenta de análise
de sentimentos na obtenção de fatores que provocavam sentimentos negativos no
público online considerando os parques nacionais individualmente. A análise, realizada
separadamente para cada parque, captou com sucesso, as percepções negativas dos
usuários do Twitter sobre cinco diferentes parques nacionais, mesmo diante de suas

104
características únicas e diferentes contextos de gestão (Capítulo 2, Fig. 5). Destacam-se
insatisfações relacionadas a implementação de regramentos, bem como tentativas de
flexibilização política, como a abertura de estradas dentro de parques e reduções do
tamanho dessas áreas protegidas.
Contudo, a análise de dados de redes sociais, requer cautela devido a desafios
relacionados à subjetividade textual como coloquialismos, gírias, ironias e questões
éticas relacionadas a coleta de dados. Ademais, é fundamental enfatizar que existe a
necessidade de aprimoramento dos conjuntos de dados para classificação dos
sentimentos na língua portuguesa. Neste trabalho foram encontradas dificuldades em
distinguir tweets positivos e neutros dos tweets negativos. No entanto, ao estender o uso
da análise de sentimentos para compreender as percepções das pessoas sobre a
conservação ambiental na língua portuguesa, esta pesquisa abre novas possibilidades e
fortalece o uso dessa abordagem metodológica para outros idiomas além do inglês. Outro
desafio a ser considerado é a representatividade das opiniões online em comparação
com as opiniões do mundo real. Mesmo diante do fato de que aproximadamente 83,6%
da população brasileira utiliza alguma rede social, é de suma importância avaliar de
maneira crítica se os resultados das percepções oriundas das redes sociais são
representativos dos usuários das áreas protegidas ou da sociedade em geral.
Reconhecendo estes desafios e o potencial inerente ao uso das redes sociais
como ferramenta de investigação, é aconselhável realizar uma limpeza criteriosa e uma
análise crítica de todos os dados utilizados, visando minimizar as possíveis
subjetividades textuais existentes. Para fins éticos, embora o conteúdo gerado pelos
usuários das redes sociais ser disponibilizado gratuitamente por algumas plataformas, é
de extrema importância considerar a privacidade e o bem-estar dos usuários. Para tal, o
compartilhamento dos dados coletados deve ser realizado utilizando critérios que
garantam o anonimato e a proteção à privacidade dos usuários. Com o intuito de
aprimorar a precisão dos modelos de aprendizado de máquina voltados à classificação
de sentimentos, recomenda-se que futuras pesquisas utilizem bancos de dados com
sentimentos previamente classificados, especialmente focados em temas ambientais.
Por fim, para mitigar os desafios relacionados à representatividade, é sugerido integrar
os resultados das pesquisas em redes sociais a diversas fontes online, como Wikipedia,

105
Instagram e Facebook. Além disso, validar os resultados online por meio de estudos de
campo amplia o potencial de representatividade dos dados, conectando as percepções
online às percepções offline relacionadas às áreas protegidas.
Em geral, ao enfrentar esses desafios, a análise de dados das redes sociais, pode
se tornar uma ferramenta valiosa na gestão sustentável e eficaz das áreas protegidas.
No contexto prático da gestão da biodiversidade brasileira, a utilização das redes sociais
pode auxiliar nas seguintes estratégias de gestão:
i.

Complementar os monitoramentos anuais já conduzidos pelos órgãos
gestores como o Sistema de Análise e Monitoramento de Gestão (SAMGE),
direcionado a avaliação da efetividade das APs, bem como o
monitoramento anual da visitação das áreas protegidas. Incluir as
percepções e atitudes da população aos sistemas de avaliação e
monitoramento

podem

proporcionar

insights

holísticos

sobre

necessidades de melhorias das áreas protegidas;
ii.

Subsidiar os planos de comunicação das APs, identificando as áreas que
exigem maior atenção da equipe gestora. Isso pode envolver uma melhor
comunicação sobre a valorização de locais de visitação significativos para
os visitantes, bem como o esclarecimento sobre os objetivos de criação,
regramentos e até possíveis problemas e conflitos para a sociedade;

iii.

Monitorar eventos que geram mais interesse público e sentimentos
negativos, possibilitando uma ação rápida da administração da área
protegida em relação a crises e eventos adversos;

iv.

Auxiliar no desenvolvimento ou revisão dos planos de manejo, planos de
uso público e de interpretação ambiental, nos quais as percepções e
sentimentos do público online podem complementar as informações e
opiniões da comunidade local em relação ao uso do espaço público
protegido.

Além dos resultados e sugestões apresentados, esta pesquisa suscitou duas
questões que merecem uma atenção mais aprofundada em futuras investigações. A
primeira diz respeito à representatividade das percepções online sobre áreas protegidas.

106
Ou seja, o quão representativos são os dados extraídos de redes sociais e até que nível
podemos considerar estas percepções como sendo a percepção da sociedade como um
todo? E a segunda está relacionada às percepções da sociedade em relação as espécies
icônicas protegidas por meio dessas áreas – dado que esta discussão, com exceção do
atropelamento de fauna silvestre, não emergiu durante a análise dos mais de 10 anos de
dados coletados sobre as áreas protegidas brasileiras no Twitter.
Neste contexto, em meio a ampliação cotidiana de ferramentas online, onde as
pessoas compartilham cada vez mais seus pensamentos, opiniões e anseios de maneira
pública e abrangente, os dados digitais das redes sociais tornam-se grandes aliados a
favor da conservação ambiental, auxiliando na identificação de lacunas e na promoção
de melhorias que conquistem o apoio do público às áreas protegidas brasileiras.

107

APÊNDICE A – MATERIAL SUPLEMENTAR
5 USING SOCIAL MEDIA AND MACHINE LEARNING TO UNDERSTAND SENTIMENTS TOWARDS BRAZILIAN NATIONAL
PARKS

Table 1: Palavras-chave utilizadas na coleta de dados no Twitter
N

Keywords used in
data collection (PT)

1 Parque Nacional
2 Parque estadual
3 Parque natural municipal
4 Parque municipal
5 Estação ecológica
6 Reserva biológica
7 Monumento natural
8 Refúgio da vida silvestre
9 Reserva extrativista
10 Área de proteção ambiental
11 Floresta nacional
12 Floresta estadual
13 Floresta municipal
14 Reserva de desenvolvimento sustentável
15 Área de relevante interesse
16 Reserva Particular do Patrimônio Natural
17 Unidade de conservação
18 Area protegida

Keywords used in
data collection (EN)

Brazilian PA category
(BRASIL, Law No 9.985, of 18 july 2000)

National park
State park
Municipal natural park
Municipal park
Ecological station
Biological reserve
Natural monument
Wildlife refuge
Extractive reserve
Environmental protection area
National forest
State forest
Municipal forest
Sustainable development reserve
Area of relevant interest
Private natural heritage reserve
Conservation unit
Protected area

Parque
Parque
Parque
Parque
Estação Ecológica
Reserva Biológica
Monumento Natural
Refúgio de Vida Silvestre
Reserva Extrativista
Área de Proteção Ambiental
Floresta
Floresta
Floresta
Reserva de Desenvolvimento Sustentável
Área de Relevante Interesse Ecológico
Reserva Particular do Patrimônio Natural
-

108

APÊNDICE B – MATERIAL SUPLEMENTAR
5 USING SOCIAL MEDIA AND MACHINE LEARNING TO UNDERSTAND NEGATIVE
SENTIMENTS TOWARDS BRAZILIAN NATIONAL PARKS

Material suplementar: Informações metodológicas sobre a análise de sentimento com
dados de tweets em português.

In the scope of our study, an extensive dataset comprising over 100,000 tweets
concerning the 74 Brazilian National Parks was acquired. The primary objective was to
discern the sentiment embedded within these tweets and ascertain the principal themes
associated with each sentiment classification. The initial phase encompassed the
classification of these tweets based on their sentiment through the utilization of a pretrained transformer model - BERTimbau Base (a.k.a. "bert-base-portuguese-cased")
(Souza et al. 2020). This model was subsequently refined to cater specifically to our task,
which involves the categorization of tweets into either positive, negative, or neutral
sentiments.
Given the focus on understanding and classifying the sentiment expressed within
tweets pertaining to Brazilian national parks, the model necessitated fine-tuning utilizing
a pre-structured dataset that had already been categorized across a spectrum of
sentiment types. Nonetheless, it is noteworthy that pre-trained models such as
BERTimbau Base are primarily designed for three fundamental natural language
processing (NLP) tasks: Named Entity Recognition, Sentence Textual Similarity, and
Recognizing Textual Entailment, and their direct application to text classification is not
inherently straightforward (Tunstall L. et al., 2022). Hence, adaptations were implemented
to align the model with our distinct classification objectives.
In this endeavor, we harnessed a comprehensive dataset originating from the B2W
e-commerce company, in conjunction with the Corpus Buscapé (B2W-Reviews01, 2018;
Corpus Buscapé, 2013). This publicly accessible dataset encompasses in excess of
130,000 user reviews encompassing a diverse range of products. Among its attributes, it

109
features binary labels denoting whether users would recommend the product to others
and ratings scored on a scale from 1 to 5 stars. It is pertinent to mention that our analysis
solely engaged with the user rating component.
Opinando, an institution specializing in mining opinions from Portuguese textual
content and established under the auspices of the Research Office of the University of
São Paulo (USP), has contributed significantly to the domain of natural language
processing (NLP). One of its notable contributions is the creation of various robust corpora
germane to this field. Among these, the Corpus Buscapé emerges as a substantial
compilation of Portuguese product reviews, harvested in the year 2013 from the Buscapé
website, renowned for its product and price exploration capabilities. Unlike the previously
referenced datasets, this corpus employs a rating spectrum spanning from 0 to 5. As a
corollary, comments affiliated with a rating of zero were omitted from the analytical
considerations.
The decision to select this particular dataset hinged on its capacity to yield precise
predictions for the classification of the target tweets. Notably, this opinion-oriented dataset
is stratified based on a "rating" feature, encompassing scores ranging from 1 to 5. A rating
of 1 signifies a markedly negative sentiment, while a rating of 5 corresponds to a highly
positive sentiment. As part of our classification strategy, ratings of 1 and 2 were grouped
as "negative," a rating of 3 as "neutral," and ratings of 4 and 5 as "positive." With the model
having undergone training and validation using this dataset, its application was extended
to a subset of approximately 2,000 tweets culled from our larger collection. This selected
subset had been manually categorized in advance to serve as a dedicated test set for
evaluation purposes.

1. How Long Are Our Tweets?
Transformer models possess a defined upper limit for the length of input
sequences, commonly denoted as the "maximum context size" or "tokens." Tokens within
these models can encompass complete words, subword components, or even singular
characters, such as punctuation marks.

110
For instance, a tweet is constrained to a maximum of 280 characters. Considering
the average of approximately 6 characters per word in the Portuguese language, the
typical tweet may contain approximately 47 words, contingent on the nature of the content
under analysis. The e-commerce sales platform’s dataset on product reviews presents
the following word frequency:

Fig. 1: Word frequency about product reviews on Buscapé.

By type of sentiment:

Fig. 2: Boxplot illustrating the sentiment per word in the reviews from the Buscapé website. The
number 1 represents positive sentiment, 0 represents negative sentiment, and 2 represents
neutral sentiment.

111
Positive sentiments generally exhibit greater word length in their expressions;
however, this trend is subject to exceptions, as evidenced by the elongated tails observed
in the boxplot representations. Hence, adaptations were implemented to align the model
with our distinct classification objectives through Transfer Learning techniques.

2. BERTopic analysis
BERTopic is a Python library for natural language processing topic modelling that
combines transformer embeddings with clustering algorithms to identify topics in a corpus
of texts (Grootendorst, 2022). The BERTopic model supports over 50 languages and has
been compared to other models, such as LDA, for performing topic modelling on short
texts from social media platforms and has shown exceptional performance in extracting
topic representations (Egger and Yu 2022).
In the first step of the BERTopic algorithm, we obtain embeddings for all documents
in the corpus, which are numeric vector representations of the documents. The next step
is to perform clustering on the embedded documents, for which dimensionality reduction
techniques, such as Uniform Manifold Approximation and Projection (UMAP), are
employed to reduce the high dimensionality of the embedding vectors (McInnes et al.,
2018). The UMAP algorithm is used by default because it preserves both the local and
global structure of the data with superior runtime performance, an important factor in
representing the semantics of text data. Preprocessing of text data is an optional step in
natural language processing. Generally, It is not recommended to remove stop words as
a preprocessing step when using the BERTopic model because transformer-based
embedding models, which we utilise, require the complete context to generate accurate
embeddings.
The default clustering algorithm used by BERTopic is HDBSCAN, which is a
density-based model that automatically identifies the number of clusters without requiring
a pre-specified number of clusters. HDBSCAN is a hierarchical density-based clustering
algorithm proposed by (Campello et al., 2013). In this algorithm, documents with higher
similarity are grouped into clusters based on cluster stability. One important characteristic
of HDBSCAN is that it does not force the assignment of a data point to a specific cluster.

112
If the data point does not fit into any similarity-based group, it is considered an outlier
(Capellaro, 2021). Once the documents are assigned to clusters, the next step is to obtain
the topic representation for each cluster using class-based Term Frequency-Inverse
Document Frequency (c-TF-IDF).
This method selects the top words with the highest c-TF-IDF scores within a cluster
to represent each topic (Grootendorst, 2022). This means that the higher the value of a
term, the more representative it is of its topic. Furthermore, following the identification of
the values associated with each term within the topics, a comprehensive evaluation and
inspection of the topics was conducted to detect any potential content that might be
misconstrued as a singular topic (See Table 1 in results). This consideration, as noted by
Egger and Yu (2022), highlights a potential limitation of the model, particularly when
dealing with extensive amounts of data for analysis. Given that the use of BERTopic also
requires significant effort due to the dynamic nature of topic structures, which change
when researchers experiment with different numbers of topics, it can be considered a
laborious task to access the topics that best represent the database. Although BERTopic
offers the advantage of leveraging domain-specific knowledge to search for specific
topics, as done in this study, this process can still be considered exhaustive.
For the purpose of our study, two main steps were undertaken: (i) identification of
potential negative topics within our corpus, encompassing all Brazilian national parks, and
(ii) segregation of tweets specifically related to the six most frequently visited parks
(ICMBio, 2021), followed by clustering to discern the prominent negative topics associated
with each individual park. To achieve this, we performed the BERTopic model with the
following hyperparameters:
● For the UMAP algorithm we set n_neighbors or the number of samples used
during the manifold approximation to 15, n_components or the dimensionality that holds
the most information possible to 5, min_dist to 0, in order to get more clustered
embeddings and selected the cosine metric to compute distances in high dimensional
space.
● For HDBSCAN we set the metric to euclidean in order to compute distances in
an array and prediction_data to True to be able to apply to our dataset later, not just to fit

113
the model, for all the datasets, no matter what park the tweets are from. And we set the
min_cluster_size parameter or minimum size of the clusters depending on the number of
observations we have. The purpose is to reach a reasonable number of topics and also
that they contain coherent information to know what they are talking about.
● For BERTopic we set the parameter nr_topics to auto in order to focus on the
interpretation of the topics. Besides we use the function CountVectorizer with a list of
portuguese stopwords and ngram_range between 1 and 2 n-gram words to be extracted,
and the function ClassTfidfTransformer in order to reduce the impact of the most frequent
words, also the MaximalMarginalRelevance function in order to limit the number of
duplicate words that we can find in each topic, and finally, the function
SentenceTransformer with the bert-base-portuguese-cased model in order to use the
same embedding model for the negative tweets selected as in the previous prediction
step.
After evaluating the possible topics generated by the model, we identified, based
on our knowledge, the topics that remained consistent across all generated models.

REFERENCES

Campello, R., Moulavi, D., Sander, J., 2013. Density-Based Clustering Based on
Hierarchical Density Estimates. Gold Coast.
Egger, R., Yu, J., 2022. A Topic Modeling Comparison Between LDA, NMF, Top2Vec,
and BERTopic to Demystify Twitter Posts. Frontiers in Sociology 7.
https://doi.org/10.3389/fsoc.2022.886498
Grootendorst, M., 2022. BERTopic: Neural topic modeling with a class-based TF-IDF
procedure.
ICMBio, 2021. Unidades de conservação federais atingem novo recorde de visitação em
2021. Accessed on 23 April 2023. <https://www.gov.br/pt-br/noticias/viagens-eturismo/2022/04/unidades-de-conservacao-federais-atingem-novo-recorde-de-visitacaoem-2021>
McInnes, L., Healy, J., Melville J., 2018. UMAP: Uniform Manifold Approximation and
Projection for Dimension Reduction. https://arxiv.org/abs/1802.03426

114
Souza, F., Nogueira, R., Lotufo, roberto, 2020. BERTimbau: Pretrained BERT Models for
Brazilian
Portuguese,
in:
9th
Brazilian
Conference,.
pp.
403–417.
https://doi.org/https://doi.org/10.1007/978-3-030-61377-8
Tunstall L., von Werra L., Wolf T., 2022. Natural Language Processing with Transformers.
O’Reilly Media, Inc.